Hybrid Local Route Generation Combining Perception and a Precise Map for
Autonomous Cars
KICHUN JO 1, (Member, IEEE), MINCHUL LEE 2, (Student Member, IEEE),
WONTEAK LIM 2, (Student Member, IEEE), AND MYOUNGHO SUNWOO 2, (Member, IEEE)
1Department of Smart Vehicle Engineering, Konkuk University, Seoul 05029, South Korea 2Department of Automotive Engineering, Hanyang University, Seoul 04763, South Korea
Corresponding author: Kichun Jo ([email protected])
This work was supported in part by the BK21 Plus Program under the Ministry of Education, South Korea, under Grant 22A20130000045, in part by the Industrial Strategy Technology Development Program under Grant 10039673, Grant 10060068, Grant 10042633, Grant 10079961, and Grant 10080284, in part by the International Collaborative Research and Development Program under the Ministry of Trade, Industry and Energy (MOTIE) South Korea, under Grant N0001992, and in part by the National Research Foundation of Korea (NRF) grant funded by the Korean Government (MEST) under Grant 2011-0017495.
ABSTRACT Autonomous cars are currently widely studied in the automotive and robotics industries because autonomous driving can satisfy the needs of human drivers regarding safety, efficiency, and comfortable driving. Behavior and motion planning for an autonomous car is one of the main requirements of autonomous driving. For autonomous cars, a local route along with its road geometry and attributes of the area is required. There are two ways to obtain the local route: perception-based local route (PBLR) and precise map-based local route (MBLR) methods. First, this paper analyzes the characteristic of both the PBLR and MBLR inference methods. Then based on this analysis, this paper proposes a hybrid local route generation algorithm that chooses the best local route between the PBLR and MBLR options, according to the perceived performance and the precise map availability. To effectively create an expensive precise map, a mapping region classification algorithm is presented to selectively choose the mapping area, where the precise map must be constructed for the MBLR inference. The hybrid local route generation algorithm with the mapping region classification allows the area used for autonomous driving to be extended while reducing the cost due to the precise map. The advantages of the proposed algorithm were verified with experiments in real traffic conditions.
INDEX TERMS Hybrid local route (HLR), perception-based local route (PBLR), precise map-based local route (MBLR), precise map region classification, autonomous cars.
I. INTRODUCTION
An autonomous car on the road has to plan its future behav- ior and motion continuously to drive safely and efficiently.
To plan the proper motion of an autonomous car, a local route that the car should follow to reach its destination must be determined and provided to the planning system. In general, there are two ways to infer a local route in an autonomous driving system. The first way is based on perception and the other is based on precise maps.
The first way to infer a local route for autonomous driving is based on perception systems, including cameras, radar
The associate editor coordinating the review of this article and approving it for publication was Bora Onat.
and lidar. Many semi-autonomous cars (levels 2-3 for SAE automated vehicle classification) produced by several auto- motive companies (Tesla, Daimler, General Motors, and Hyundai Motor Company [1]–[4], and the many startup companies [5], [6]) can recognize local routes based on environmental characteristics such as lanes, guide rails, and preceding vehicles. This method is similar to the principle of human local route creation for normal driving, so the autonomous car does not need a precise map or localization, and instead only needs the topological map that is used by navigation systems. However, since current technologies for finding local routes based on the environment are primitive, it is difficult to infer the local route based on the percep- tion only in cognitively complex road situations such as
intersections, road splits, merging, and complex urban roads.
Therefore, current autonomous cars using perception-based local routes only operate in restricted environments such as highways.
The second approach using local route inference is based on a precise map with a precise localization system. The precise map includes pre-defined road information (such as the roadway geometry, lanes, and traffic attributes) with centimeter-level accuracy. The road information located in the vicinity of the autonomous car can be extracted from a precise map by applying a pose estimation based on pre- cise localization. The local route of an autonomous car can be inferred from the extracted road information from a predefined precise map. Since this method is indepen- dent of perception performance and can reliably infer local routes, many technical demonstrations of fully automated vehicles (SAE level 4-5), such as DARPA Grand and Urban challenges [7], [8], Google self-driving car [9], Daimler Bertha project [10], and several startup companies’ demon- strations [11], [12]. and A1 of Hanyang University [13], [14], operate based on precise map-based local route inferences.
However, local route inferences based on precise maps have the limitations of high cost and complicated system architec- ture due to the necessity for a precise map.
The important point here is that the two methods of local route inference have different pros and cons. A local route inference based on perception is affordable and simple because it does not rely on a precise map or localization.
However, perception-based local route inference works only in a well-structured area due to the limitations of current per- ception technologies. In contrast to perception-based infer- ence, a precise map-based local route inference can reliably find a local route regardless of its perceptual complexity.
However, local route inference requires costly precise maps and localization, thus increasing the system cost and system architecture complexity.
These different characteristics of the two local route infer- ences are complementary to autonomous driving on various types of roads. For instance, imagine an autonomous car that is based on a hybrid local route inference that uses a perception-based local route on the highway and utilizes a precise map-based local route in urban areas, respectively.
Since an autonomous car does not require precise maps on the highway, we can reduce the cost of creating and managing the precise map in that region. Since an autonomous car can drive in a complex urban area with a precise map-based local route, the coverage for autonomous driving can extend beyond the highway.
In order to implement this idea, this paper proposes a hybrid local route (HLR) generation algorithm that deter- mines a proper local route from the perception-based local route (PBLR) or precise map-based local route (MBLR) according to the precise map availability. In order to properly select one local route between PBLR and MBLR, the map- ping region classification for precise map building is a key factor, which determines the precise map availability. Based
on an evaluation of perception performance for real time local route inferences, the classification algorithm classifies the mapping region as to whether a precise map is needed or not. If the perception system can reliably detect the local route in real time, a precise map is not required; however, if the perception system has difficulty detecting the PBLR, a precise map for MBLR inference is required in that area. By applying the proposed algorithms to the autonomous driving system, it is possible to extend the coverage region needed for autonomous driving while reducing the mapping cost by reducing the mapping area. The validity of these advantages of the proposed algorithm was verified through intensive experiments in real traffic situations.
This paper is organized as follows. Section II presents the definition of a local route as well as the inference meth- ods. Section III describes an HLR generation algorithm, and Section IV describes the mapping region classifier of the precise map. Section V provides verification results based on experiences. The final section includes the conclusion and discussion of future work.
II. LOCAL ROUTE INFERENCE FOR AUTONOMOUS DRIVING
A. DEFINITION OF LOCAL ROUTE
The goal of an autonomous car is to reach its destination automatically, taking into account the various requirements of passengers such as safety, comfort, time, and energy.
To drive an autonomous car with the above requirements, the autonomous driving system must continuously plan its future behavior and motion. For planning and control of an autonomous car, an environment model that contains all of the information about the surrounding environments and roads should be provided to the planning system.
An environmental model can be generated from online perceptions, wireless communications, and predefined maps.
Components of the environment model can be classified into three parts: objects, traffic control devices, and routes.
An object is something that has the potential to collide with the autonomous car. Types of objects can be divided into static objects (buildings, walls, trees, poles, and barriers) and dynamic objects (pedestrians, bicycles, motorcycles, and vehicles). The traffic control devices, including road surface markings, speed bumps, traffic signs, and traffic lights, pro- vide information about traffic rules that must be followed at that place and moment. The behavior planning of autonomous cars must take into account the information from traffic con- trol devices to avoid accidents with other road users and to facilitate the flow of traffic.
The last part of the environment model is the route, which represents the paths or roads that the autonomous vehicle should follow in order to reach the destination. The route can be categorized as a global route or a local route in accor- dance with the defined coordinate system and the intended use of autonomous driving. The global route is defined in a global coordinate system (world geo-reference coordinate), whereas the local route is defined in a local coordinate system
FIGURE 1. Global route and local route for autonomous driving.
FIGURE 2. Definition of the local route that is one of part of an environment model.
(vehicle body coordinate). The global route is a series of road networks linked from a departure point to a destination. This is generated using a global routing algorithm that optimizes various costs, such as travel distance, time, and energy, just like a commonly used GPS navigational system for human drivers.
The local route, which is a part of the global route around the autonomous car (Fig. 1), aims to provide the following roads and boundary information to the autonomous car at that time. In addition, the local route should be updated in real time by taking into account the execution period of the planning or control approach. Although the local route includes important information for autonomous driving, there is no specific definition of a local route in previous research on autonomous driving. This is because previous studies did not define the local route as an independent data structure for their environmental model. They used the local route as an internal part of the implementation of perception, planning, or the control algorithm. To clarify the roles of local routes when driving autonomously, this paper defines a specification for the local route (see Fig. 2) as the following.
1) Geometry: the local route provides geometric infor- mation that must be followed by the autonomous car. The geometry guides the vehicle to the final destination, but it does not need to provide all of the geometry from the vehicle location to the destination. For the application of planning and control, the length of the geometric information should be sufficient to be accessible within a reasonable time by autonomous vehicles. The geometry information can be rep- resented by several data structures, such as discrete waypoints (sequences of points) and various types of continuous curves (polynomial, Clothoid, and spline).
2) Boundary: the autonomous car must go along the geometry of the local route, but this does not imply that the local route geometry is the best path for autonomous driv- ing since there is no consideration of real-time cost factors such as vehicle dynamics, surrounding objects, and traffic control devices. A motion planning algorithm can generate a smoother optimal trajectory within a certain range by taking into account real-time factors. This range can be defined as the boundary of the local route, and the inside of the boundary must be drivable space, as shown in Fig. 2. The boundary information can also be represented by discrete points or various curves that can be represented by a geometric data structure of the route.
3) Vehicle-centric coordinate system: the motion- planning algorithm generates the desired trajectory of an autonomous car based on the current position of the vehicle center. The vehicle controller controls the vehicle motion for the trajectory tracking based on the vehicle motion models (such as a kinematic and dynamic models) represented on the vehicle-centric coordinate system. It is therefore suitable for planning and controller to represent the local route with the same coordinate system. The Cartesian coordinate system is most widely used in the vehicle-centric coordinate system.
A curvilinear coordinate system can be used for the local route by using geometry as a base axis [15], [16].
4)Transition:For instance, suppose a car drives on a road with multiple lanes. The car can change the lane to the left or right to reach the final destination or to avoid a slow pre- ceding vehicle. At this time, the driver must know the states of the left and right lanes correctly. As with this example, the behavior or motion planning algorithm for autonomous cars can change from the current local route to other local routes in order to follow a global route or to achieve a desired speed. To change the local route, the planning algorithm must have information about alternate local routes.
5)Real-time: Planning and control must be executed in real-time to react to changes in the surrounding environments.
Therefore, the local route must also be provided in real time so it can be planned and controlled.
We define the local route with five features; however, detailed implementations of the local route can be different depending on the configuration of the autonomous driving system.
The local route can be extracted from two components of the autonomous car: perception and a precise map. The perception portion interprets the environment and extracts the local route directly. From the precise map, the local route can be extracted based on the position of the vehicle. These two methods for local route inference will be covered in more detail in the following subsections.
B. PERCEPTION-BASED LOCAL ROUTE (PBLR) INFERENCE The perception system of the autonomous car can interpret data from different sensors such as cameras, radar, and lidar in order to understand and model the surrounding environments.
The perception approach can infer a local route based on its
understanding of the surrounding environment. For instance, the local route can be inferred real-time perception of lanes, road boundaries, drivable spaces, and the movements of other vehicles using the vehicle onboard sensors. Especially, the lanes on roadways are good landmarks for inferring the local route because they can be detected by cameras relatively easily due to the clear difference in intensity compared to the road surface. Besides, a lane directly represents guide information for the vehicle in the lane. Therefore, many types of automated driving assistance systems (or the automated vehicles of SAE level 2), such as lane keeping systems, highway driving assists, and traffic jam assists, are based on lane PBLR inferences [17]–[20]. In addition to camera- based lane detection, the fusion of other perceptions, includ- ing guide rails, drivable spaces, and proceeding vehicles, is able to handle temporary loss of the lane due to unpredicted disturbances [21]–[32].
The advantages of PBLR inference come from its simple, independent structure. Since the local route can be inferred directly from the perception system, an autonomous driving system does not need to depend on other system compo- nents, such as a precise map or localization. If the perception approach can reliably infer the local route, the autonomous car can drive on all types of roads without a precise map. This independence enables the development of a cost-effective autonomous driving system.
Humans also use PBLR inference for driving, so human drivers do not need a precise map or localization system for driving. Based on sensory information from the eyes, ears, and sense of inertia, the human drivers instantly inter- pret their surrounding environments, find a local route, plan their trajectory, and control the vehicle by using the steering wheel, throttle, and brake. If the perception sys- tem on the autonomous car has human-level perception, the autonomous car can go anywhere with only a navigation map.
Unfortunately, the current technology of machine per- ception is less capable than a human in terms of sensors, algorithms, and computation. Therefore, PBLR inference is only available in limited areas, such as highways and well-structured roads. The current technologies for percep- tion are limited in their ability to infer a local route in perceptually complex conditions, including complex urban areas, unstructured roads (off-road and rural road), transitions of roads (intersections, roundabouts, exits, and merges), and special structure roads (construction site and tollgate). For this reason, the PBLR applies only to the semi-automated vehicles of SAE level 2 or 3.
C. PRECISE MAP-BASED LOCAL ROUTE (MBLR) INFERENCE
Although PBLR inference capability is rapidly improving due to recent research on machine learning algorithms (such as a deep learning), there is still a large gap between machines and humans in real driving situations. To fill the gap, a precise map with precise localization can be applied for local route
inference. The precise map includes the roadway geometry and attribute information at the centimeter-level [33], [34].
Since the mapping is done off-line, it is available to improve the accuracy of mapping by applying a powerful algorithm that requires significant computational power and databases.
The local route near the vehicle can be determined by search- ing the roadway and attributes from the precise map using pose estimation for precise localization.
Many previous demonstrations for fully autonomous driv- ing utilized the MBLR inference for planning and con- trol. In the DARPA’s Grand Challenge competition, one key technique is managing the waypoint map to operate the autonomous car [7]. In DARPA’s Urban Challenge, the Route Network Definition File (RNDF) that contains the road- way geometry (waypoints) and attributes (stop sign, lane widths, and checkpoints) was mainly used to conduct the planning and control of the autonomous cars [8]. Google’s self-driving cars utilized a highly accurate three-dimensional map for autonomous driving [9]. In Daimler’s demonstration on the Bertha-Benz-Memorial-Route, Lanelets (an efficient map representation for drivable environments) was applied to the highly precise map used for autonomous driving [35].
The author’s demonstration, the autonomous car A1, was also based on the B-spline roadway geometry map and road surface markings (RSMs) with centimeter-level accu- racy [13], [14], [33]. In addition, many map companies, such as HERE and TomTom, provide the High Definition (HD) map for the autonomous driving [36], [37].
The advantage of the MBRL inference is due to the robust and stable performance of recognition of the local route.
Since the roadway geometry and attributes are already stored in the precise map, the local route can be reliably extracted from the known geometry based on precise localization with a simple calculation process. Therefore, the autonomous driv- ing system can reliably infer the local route in perceptu- ally challenging environments (such as complex urban areas) based on the MBLR inference. Because of this robustness, many demonstrations for level 4-5 automated driving were utilized based on MBLR.
However, a disadvantage of MBLR inference is its depen- dence on the precise map and localization. The autonomous cars using the MBLR are only available to drive in the regions that have the precise map. To construct the precise map for a wide area requires significant data acquisition mapping and continuous map management. In addition, the complexity of the autonomous driving system is increased due to the use of precise localization and the precise map database.
In conclusion, MBLR inference is a more expensive solution than PBLR inference.
III. HYBRID LOCAL ROUTE BASED AUTONOMOUS DRIVING
A. COMPLEMENTARY NATURE OF PBLR AND MBLR Table 1 summarizes the characteristics of PBLR and MBLR, respectively. In terms of the system complexity and dependency, PBLR inference has advantages over MBLR.
TABLE 1. Differences between PBLR and MBLR.
TABLE 2. Characteristics of hybrid local route (HLR) inference.
Since PBLR infers the local route based only on its own sensor sensors, the autonomous driving system based on a PBLR can have a simple system architecture and indepen- dence from other factors (such as a precise map and local- ization). However, since MBLR requires a precise map and localization, the autonomous car must take into account these systems in its system architecture. Especially, a precise local- ization system requires several sub-components, such as vehi- cle motion sensors, GPS, and landmark alignment systems, in order to obtain centimeter-level accuracy of the vehicle pose.
The high complexity and dependency of MBLR inference increase the cost of the entire autonomous driving system.
Above all, a precise map has high costs due to the process of map construction and management. Many mapping sys- tems are based on a survey using mobile mapping vehicles.
Therefore, in order to generate a precise map in certain areas, the mobile mapping vehicles explore the area to acquire a precise map. After surveying and acquisition, the roadway geometry and environmental attributes are extracted from the acquisition data by costly processes. In addition, network and database systems are necessary to share the large data volume of the precise map with the autonomous car. These costs will increase in proportion to the size of the mapping area. In addition to the precise map, the costs of the precise localization for localization computing and related sensors are required.
However, although the cost of MBLR inference is high, the autonomous driving system cannot rely solely on PBLR because of the superior performance of MBLR in com- plex environments. Since the MBLR can be extracted from a predefined precise map, it can reliably find the local route without perceptual limitations such as being blocked by other objects, limited perception range, and recognition performance.
FIGURE 3. Local route inferences based on perception and precise map.
Based on a comparison analysis of Table 1, the two local route inferences are complementary. Complementary routes support hybrid local route (HLR) inferences, which adap- tively select the one local route according to the perceived driving situation. If the driving conditions are simple enough to infer the local route reliably using perception (such as the highway), the autonomous car selects the PBLR for autonomous driving. Since the autonomous car can be oper- ated in these regions only based on perception, a precise map is not required in those regions. In perceptually complex sit- uations (such as urban ones), the autonomous car switches to the MBLR inference, and precise maps should be constructed for those regions before autonomous driving.
We can expect the characteristics of the HLR inference as described in Table 2. Although the system complexity increases due to the hybrid mechanism, mapping costs and precision map requirements applying the HLR strategy to an autonomous driving system make it possible to expand the autonomous driving area at an effective cost.
B. HYBRID LOCAL ROUTE (HLR) GENERATION BASED ON THE SELECTION BETWEEN PBLR AND MBLR
The HLR is generated by a selection between PBLR and MBLR, as shown in Fig. 3. The PBLR is generated using perception information (such as lanes and road boundaries) coming from the various perception sensors. The MBLR is extracted from the roadway geometry and attributes of the precise map by applying the pose estimation of the precise localization system. For precise localization, various types of sensors (such as GPS, vehicle motion sensors, and per- ceptions) are used to estimate the vehicle pose. The selected local route is used for planning and control of the autonomous driving system.
The selection algorithm simply chooses a local route between PBLR and MBLR based on the availability of a
precise map for the driving region. In other words, if the precise map is available in the driving region, the algorithm selects MBLR; otherwise, it selects PBLR. The availability of the precise map for the driving region are determined based on a mapping region classification algorithm at the precise map generation stage. Therefore, the process of mapping the region classification for building precise maps is a key factor in HLR generation. Section IV will present the map- ping region classification to distinguish where the precise map should be constructed due to the limitations of PBLR inference.
IV. MAPPING REGION CLASSIFICATION OF PRECISE MAP A. PRECISE MAP FOR THE AUTONOMOUS DRIVING A precise map for autonomous cars contains detailed infor- mation about the physical world, which can be used by the autonomous driving system. This information about the phys- ical world on the precise map can be represented by various types of features and their attributes. The examples of the features on the precise map are the physical states (such as position, attitude, and size) of the roadway geometries, road surface markings (RSM), traffic control devices, and road boundaries. Examples of such attributes are the number of lanes, speed limits, and road types. Unlike the general navi- gation map, the features on the precise map are very dense and accurate (centimeter level) and the attributes contain more specific information about the features. For example, whereas the navigation map contains topological information on the lane features (no concern about the number of lanes) to assist the driver, the precise map contains detailed marking infor- mation for each lane at the centimeter level. These precise features and attributes on the precise map can be applied to the autonomous driving system to improve the perception coverage or identify landmarks.
In order to construct a precise map for autonomous driv- ing, generally three steps are required: data acquisition, data processing, and database management. The first step is data acquisition, which includes collecting information about the physical world using a mobile mapping vehicle equipped with several mapping sensors such as GPS receivers, inertial sen- sors, cameras, and lidar. The second step is the data process- ing, which extracts the feature and attribute information from the collected data of the previous step. Many types of feature extraction systems or manual extraction processes are needed to distinguish the features and attributes. The final step is the construction of a map database to allow autonomous cars to gain access to the features and attributes. Data storage and network service such as a cloud service are necessary to manage the precise map database.
The construction cost of the precise map database is much higher than the cost of the common topological navigation map. The reasons for the high costs are that the mobile map- ping vehicle should drive through all of the mapping areas to obtain the raw data, the feature and attributes extraction processes need huge processing times, and the costs of the data storage and network service are not affordable. However,
if we can selectively generate the precise map for the specific area where the MBLR inference is necessary (or where the PBLR inference cannot cover), the construction cost of the precise map can be reduced (we can reduce the second and third process). Therefore, this paper proposes a classification algorithm of the mapping region, which selectively generates the precise map based on the performance evaluation of PBLR inference.
B. MAPPING REGION CLASSIFICATION
The classification process for the mapping region of the precise map can be added after the first step (data acquisition) of the three-step construction process for the precise map. The first step of data acquisition is obtaining the raw data for the driving area of interest; however, we can reduce the mapping cost for the second step (feature and attribute extraction) and third step (database and network) of the mapping process, which are the most expensive tasks, by selectively classifying the mapping area.
The mapping region classification for the precise map is performed based on a performance evaluation of the PBLR inference. If the perception is stable enough to recognize the exact local route in some region, the precise map is not required for that region because the autonomous car can find the local route without a precise map. This means that the classifier should select the areas with poor performance in terms of PBLR inference for the mapping region of the precise map.
The mapping classification process can be divided into three steps: 1) a reference route extraction, 2) a PBLR evalu- ation, and 3) a mapping area classification as shown in Fig. 4.
1) REFERENCE ROUTE EXTRACTION
At the first step, the reference route can be extracted from the driving trajectory of the mobile mapping vehicle, as shown in the blue curve of Fig. 4 (step 1). Since the mobile mapping vehicle follows the route identified by professional drivers for mapping data acquisition, the trajectory of the mobile mapping vehicle can be a reference route to evaluate the PBLR inference. In order to obtain more accurate and reliable driving trajectories, it is possible to apply the post-processing process of the positioning system [33] or many different types of offline processing methods (such as GraphSLAM or trajectory optimization) depending on the mapping system configuration.
2) PBLR EVALUATION
PBLR performance can be assessed by analyzing the geo- metric consistency (similarity) of the PBLR relative to the reference route. The simplest way to evaluate the degree of consistency for two routes is to compute the area between two route geometric models (PBLR and reference route), as shown in the Fig. 5b. However, only area evaluation is not enough to represent the consistency because of two reasons.
First, as the length of route increases, the area increases.
Since a long route pair has larger area than a short route
FIGURE 4. Mapping classification process: a reference route extraction and PBLR evaluation.
pair with same shape, we cannot evaluate the consistency of route pairs with different lengths. To normalize the length effects, the area is divided by an arc-length of reference route, and this value is same as an average of error distance between the two routes for the reference route arc-length.
Second, each part of the route has a different importance.
For instance, the route part near the autonomous car is more important than the far away because the close part is directly used for motion planning and control of the autonomous vehicle. The importance of the route part can be modeled in weight function on the arc-length of reference route by taking into account the motion planning and control characteristics, as shown in the Fig. 5a. The consistency index with respect to both weight and length is obtained by dividing the weighted area by the sum of the weight. This consistency index value represents the weighted average of the error distance between the two routes for the arc-length direction. If the two route geometric models are analytically integrable for the reference route arc-length, we can find an analytic equation to get the consistency index. However, since the mathematical form of the route geometries are not specified and vary depending on the PBLR system, this paper proposes a consistency distance function based on sampling-based error evaluation, which is an approximation of weighted average of error distance between two routes. This approximation guarantees the gen- erality of mathematical form of routes and the computing efficiency.
Overall process of the consistency distance function based on sampling of error distances is shown in Fig. 6. The pro- posed consistency distance function computes the consis- tency distance between the reference route geometry and the
FIGURE 5. Consistency index for two different geometries: weighted average of the error distance.
FIGURE 6. Consistency distance function to evaluate the geometric consistency between two different routes.
PBLR geometry with its importance weight considering the characteristics of the autonomous driving functions. Depend- ing on the local route format and the PBLR inference perfor- mance, the boundary and transition local routes of PBLR can also be added to the calculation of the consistency distance.
Computing process of the consistency distance function consists of four steps. The first step is a sampling of pointsli
from the reference route geometry at a regular interval (Fig. 6 (a)). When determining the regular interval value, we must consider the geometric complexity of the route and the com- puting efficiency. If the regular interval is dense, the func- tion can accurately evaluate the consistency for two routes even with complicated geometries, but the computational load increase. In contrast, if the regular interval is sparse, the computing speed will be improved, but the consistency assessment using the function is not accurate for the com- plicated route geometry. The initial sample point l0 should be the closest one from the center of the autonomous car in order to take into account the consistency weight of the third step. The second step is calculating the shortest Euclidean
distancese(li) from the sample points of the previous step to the target geometry (Fig. 6 (a)). The third step is to generate the consistency weightw(l) which represents the importance of the calculated distancee(li) regarding the effective distance (Fig. 6 (b)). The effective distance represents the valid range for a local route that is actually used by the planning or control algorithm. The consistency weight is modeled based on the modified Cumulative Distribution Function (CDF) of a Gaus- sian probabilistic distribution with N(leff, σeff2 ), where leff means the effective distance andσeff represents the standard deviation of the effective distance uncertainty, as described in an equation (1).
w(l)= 1 σeff
√ 2π
Z l
−∞
e
−(−l+leff)2
2σ2
eff dl (1)
The sample’s distances located inside the effective distance have valid weights, and the distance samples located outside the effective distance are not valid for the weighting. The uncertainty of the effective distance can determine the gra- dient of the weight transition between valid and invalid. The final step (Fig. 6 (c)) is to calculate the consistency distance by applying a weighted average of the distancee(li) for the consistency weightw(l), using an equation TABLE2. Since the consistency distance function is a weighted average of sample distance errors, the output value is scalar with meter unit.
Distconsistency=
N
P
i=1
w(li)e(li)
N
P
i=1
w(li)
(2)
3) MAPPING AREA CLASSIFICATION
The mapping area for a precise map can be classified through the consecutive threshold of the consistency distance between the PBLR and the reference route. If the PBLR can provide proper local routes that have a consistency distance less than the evaluation threshold Diste_t with the specific number Ne of consecutive steps, then this region is classified as a non-mapping area, as shown in the green color curve of Fig. 4 (step 3). However, if the consistency distance for the PBLR is larger than the evaluation thresholdDiste_t, this region is classified as a mapping area for the precise map, as shown in the red color curve of Fig. 4 (step 3). The conservativeness of the mapping area selection can be determined based on the adjustment of the evaluation thresholdDiste_t of the consis- tency distance and the number of consecutive stepsNe.
V. EXPERIMENTS A. TEST ENVIRONMENTS
The autonomous car A1 used to evaluate the proposed HLR generation is shown in Fig. 7. For the PBLR inference, the test car is equipped with a front lane detection camera manufac- tured by Mobileye. For the MBLR inference, the autonomous car must have a precise localization system and a precise
map. The precise localization system estimates the precise position and heading of the vehicle based on the matching of real-time landmark perception with the landmark features stored in the precise map. Based on the author’s previous localization study [38], a Monte Carlo localization algorithm was applied for the matching of perception and map features.
Road surface marking (RSM) features painted on the road surface (such as lanes, zebra crossings, and arrows) were used for the localization landmark. In order to detect the RSM features, the test vehicle had several cameras such as an around-view monitoring (AVM) system that provides a surrounding bird’s-eye view image, a front camera, and a rear camera. Wheel speed sensors and inertial motion sen- sors were installed to obtain the motion information for the vehicle. Evaluation of the HLR was performed by applying the HLR to autonomous driving. For HLR-based trajectory planning of the autonomous car, a sampling-based motion planning algorithm that was developed by the present authors was applied [39]–[41].
The test site for the evaluation of the proposed HLR gener- ation was an urban area located in Incheon, Korea, as shown in Fig. 8. The trajectory marked with a yellow line is the total test route for the evaluation, and its length is 14.65 km.
Since the test site was composed of a combination of the perceptually easy areas (clear lane marking) and difficult areas (no lane marking), it was a good place to evaluate the proposed HLR generation for autonomous driving.
B. LOCAL ROUTE SPECIFICATION AND INFERENCE The definition of the local route was described in Section II and Fig. 2. In order to satisfy the definition, we designed the local route according to the following specification.
1) The geometry of the local route is represented by a cubic-spline for a maximum of 100 meters.
2) The boundaries of the local route are also represented by cubic-splines and are located at the left and right of the geometry, respectively.
3) The geometry and boundaries of the local route are represented in the vehicle-centric coordinate system whose origin is located at the center of the vehicle rear axis.
4) The transition routes for the left and right of the center geometry can be provided when each transition exists.
5) The inference of the local route should be synchronized with an execution period for motion planning.
For PBLR inference, a front view camera (developed by Mobileye, Fig. 7) was used. The front camera can directly provide lane information for the center, left, and right of the vehicle via the CAN network. The provided lanes were converted to a B-spline to follow the geometry and boundary specifications of the local route. For synchronization with the planning algorithm, the converted geometry and boundaries were adjusted based on the perception timestamp and vehicle motion information.
For the MBLR inference, a precise map that stores the global geometries and road boundaries of roadways was used along with the RSM-based precise localization system. Since
FIGURE 7. Autonomous car A1 for evaluating the autonomous driving performance of the online selection-based local route inference.
FIGURE 8. Result of the precise map generation and mapping region classification.
the precise map included the global roadway geometries and road boundaries in the same format as the local route (cubic spline), it is straightforward to extract the MBLR from the global precise map. By using the precise vehicle pose esti- mated from the RSM-based precise localization, the MBLR around the vehicle can be extracted from the precise map by searching for the geometries and boundaries located inside the region of interest (ROI) of the planning algorithm. The extracted MBLR should be converted into the vehicle centric coordinate using the pose estimate of localization. Since the MBLR was extracted based on the pose estimate of precise localization, the synchronization with the motion planning algorithm was performed based on the precise localization timestamp.
C. PRECISE MAP GENERATION BASED ON THE MAPPING AREA CLASSIFICATION
1) PRECISE MAP FOR MBLR
A precise map was used for the MBLR inference and precise localization. For the MBLR inference, the precise map must contain geometries and boundaries of roadways in the global coordinate system, as shown in Fig. 9-(a). For precise local- ization, the precise map contained landmark features (RSM in Fig. 9-(a)) that can be aligned with the RSM perception feature in order to estimate the position and heading of the vehicle. In order to construct a database for the precise map, OpenStreetMap (OSM), a free editable map in XML format, was used due to its scalability and interchangeability. The road geometries and boundaries were described in the OSM nodes and ways, and the height, arc-length, and cubic-spline
FIGURE 9. (a) Open Street Map database for a precise map that contains the RSM, geometry, and boundary information, (b) OSM data structure for the representation of RSM, geometries, and boundaries.
FIGURE 10. Mapping system equipped with centimeter-level accuracy RTK-GPS and two down-looking cameras are used to capture the road surface image.
parameters were included as tags (RSM in Fig. 9-(b)). The RSMs were described as polylines, so a standard OSM format can be used for the RSM database.
2) MAPPING SYSTEM FOR PRECISE MAP BUILDING
A mobile mapping vehicle equipped with a special mapping system was used to build the precise map, as shown in Fig. 10.
A high accuracy positioning system (RT3002 developed by
OxTS) that processes a real-time kinematic (RTK)-GPS and inertial navigation system (INS) was used to measure motion information and a reference trajectory (position and heading) of survey route. Two down-looking cameras mounted on the tail of vehicle captured the road surface images for extracting RSMs required for the precise localization.
After collecting all the raw sensor information for an autonomous driving area using the mobile mapping vehicle, we have to apply a data processing to extract the roadway geometry, boundary, and RSMs for the precise map. First, we process the collected INS and RTK-GPS data to optimally estimate the reference trajectory using the optimal smoothing technique [33]. The optimal smoothing algorithm filtered the RTK-GPS error due to the GPS signal outage. Second, we can extract the RSMs using the image processing and manual inspection. Bird’s-eye view images converted from the images of two down-looking cameras were stitched into a single global image using the corresponding smoothed ref- erence trajectory. RSM features can be extracted by applying top-hat kernel filter to the global stitched image [41]. Among the RSM feature results, fault and undetected RSM features were manually corrected. Finally, based on the processed reference trajectory and RSM features, we can extract the roadway geometry, boundary, and RSMs into the OSM for- mat. However, this process would be a very expensive task if we perform the process for all the mapping area. In order to efficiently perform the post-processing, we applied the mapping region classification algorithm, which selectively processes the raw mapping data for the precise map building.
3) MAPPING REGION CLASSIFICATION
The main idea of mapping region classification was to iden- tify areas where PBLR inference cannot work. If the PBLR inference was not reliable enough to identify the local route, the region is classified as the precise map building region because the autonomous car can infer the MBLR local route based on a precise localization and map.
The mapping classification process consists of three steps that are a reference route extraction, a PBLR evaluation, and a mapping area classification. First, the reference route was constructed based on an optimal smoothing algorithm using RTK-GPS/INS [33]. The PBLR inference data was synchronized with the time stamp of reference route using the GPS time stamp. Second, we performed the PBLR evaluation by evaluating the consistency distance between the reference route and PBLR inference. For the consistency distance func- tion,leff (effective distance) andσeff (the standard deviation of the effective distance uncertainty) were set to 20 meters and 5 meters, respectively, by taking into account to look ahead distance of vehicle controller. Finally, the mapping classification algorithm was performed to determine the pre- cise map building region. The threshold parameters used in the classification algorithm were a consistency distance of 0.2 meters and five consecutive steps.
Fig. 11 and Table 3 show the result of mapping region clas- sification according to the variation of consistency threshold
FIGURE 11. Result of the precise map generation and mapping region classification.
(0.1 – 0.5 meters). In the figure, the red zone is classified as a mapping region for the precise map, and the green zone represents an area not classified as a mapping region.
By applying the mapping region classifier, we were able to reduce the mapping area up to 84.88 % for the 0.5 meters consistency threshold. Fig. 12-(a) represents an area where was not selected as precise map building region because it had very clear lane marking for reliable PBLR inference. In con- trary, Fig. 12-(b) represents a precise mapping area where the local routes were difficult to find through the lane perception due to the zebra crossing and speed bumps. The consistency distances for the (a) and (b) are also shown in Fig. 12. The consistency distance for the (a) region has small consistency distance (0.0099 meters), whereas the (b) region has large distance (0.5278 meters).
To assess the reduction in mapping costs due to the map- ping region classification, we compared the man-hour to construct the precise map with and without mapping region classification for the 0.3 meters consistency threshold case.
TABLE 3. Mapping region by consistency threshold.
FIGURE 12. Consistency distance calculation for (a) non-mapping and (b) mapping regions.
For extracting the landmark features (RSM) and attributes of the precise map (step 2 and step 3 for precise map generation), it requires an average of one man-hour for the 100 meters of the road using the in-house developed auto-recognition tool and manual annotations. Besides, for the complex landmark areas such as the red zone in the Fig. 12, it took two to three times as long to work. As the results of the comparison of the actual working man-hours, we were able to reduce the work- load by about 80 % using the mapping region classification algorithm.
D. AUTONOMOUS DRIVING TEST BASED ON THE ONLINE SELECTION ALGORITHM
In order to validate the applicability of HLR generation for autonomous cars, we have applied the HLR genera- tion algorithm to autonomous driving demonstration based on an autonomous driving system A1 of Hanyang Univer- sity [13], [14]. The system contained the sampling-based planning and model predictive control (MPC) based vehicle control algorithms. The motion planner generated an optimal trajectory to follow the HLR every 50 milliseconds. Based on the trajectory, the lateral and longitudinal control determined steering wheel angle, throttle position, and braking force every 10 milliseconds.
The demonstration route is shown in a map at the top of Fig. 13 as a red trajectory. This route contains urban road elements: straight and curved roads, intersections, and stop lines with crosswalks. On this road, the HLR-based autonomous driving determined its speed by considering a legal speed (60 km/h) and road curvature. Furthermore, the vehicle temporarily stopped at the front of each stop line. For the validation of the HLR algorithm, this test was
FIGURE 13. Consistency distance between the partial precise map-based HLR and the full precise map-based MBLR.
conducted on an environment without other traffic partici- pants (a vehicle, bike, bicycle, and pedestrian). The MBLR and HLR are executed on Intel i7-4700EQ CPU 2.4G com- puting unit in parallel, and the sampling period of HLR generation is 100 milliseconds.
As a result, the HLR-based autonomous driving was oper- ating without fail, and there was no significant difference from the previous autonomous driving based on a global precise map and localization in terms of driving performance (such as safety and comfort).
In addition to the qualitative performance validation, in order to quantitatively verify the performance, we eval- uated the consistency distance between the partial precise map-based HLR and the full precise map-based MBLR.
The full precise map-based MBLR can be a reference local route for the verification because it can generate the most reliable and accurate local route that is independent of the perception performance. Five plots in the Fig. 13 shows the consistency distance between the HLR and MBLR. The gray area represents where the HLR selects the MBLR, so the consistency distance is zero. The consistency distances for the partial HLR and the full MBLR are always smaller than the configured consistency threshold that can guarantee the
reliable autonomous driving. These qualitative and quanti- tative analysis results show that the autonomous driving is possible without completely including the expensive precise map for the entire driving area.
VI. CONCLUSION
This paper presents an HLR generation algorithm for deter- mining the proper local routes for PBLR and MBLR, and pro- posed a mapping region classification algorithm for efficient precise map building. The validity of the proposed algorithm was verified with applications to real autonomous driving experiments.
1) A local route that is one of the environment models for autonomous driving can be obtained from PBLR and MBLR inferences. The PBLR can be recognized by perceptions (such as lane detection), and the MBLR can be extracted from the precise map and localization.
The online selection process in the HLR generation algorithm chooses one of the local routes based on the availability of a precise map. The selected local route is provided to the planning or control functions of the autonomous car along with the information for other environment models (objects and traffic control devices).
2) The mapping region of the precise map can be classi- fied based on the performance evaluation of the PBLR inference. If the PBLR inference is evaluated as proper to recognize the local route for a certain region, the precise map for MBLR inference is not needed for the region. If the PBLR is evaluated not to find the local route due to the performance limitations for certain regions, the autonomous car requires a precise map to perform MBLR inference. The PBLR inference can be evaluated using a consistency distance function by evaluating the consistency between the reference route of the vehicle trajectory and the PBLR.
3) The HLR generation with mapping region classifica- tion provides many benefits due to the complementary natures of the PBLR and MBLR inference approaches in terms of dependency, cost, and local route perfor- mance. The precise map for MBLR inference is selec- tively built for perceptually complex environments, so the mapping cost of the precise map can be reduced.
In perceptually complex environments, an autonomous car uses MBLR inference to extend the autonomous driving area.
4) The experiments show that the mapping region clas- sification algorithm can reduce the mapping cost of the precise map. Also, autonomous driving based on the HLR generation algorithm is not different from MBLR autonomous driving, which is the most reli- able but expensive approach. This result means that autonomous driving based on HLR generation with the mapping region classification algorithm allows the autonomous driving area to be extended while reducing the cost of the precise map.
The proposed algorithm infers a proper local route based on the discrete selection of PBLR and MBLR. However, we can obtain a better local route by smoothly integrating the PBLR and MBLR approaches using information fusion.
To check the validity of the fusion approaches, the authors plan to attempt the integration of PBLR and MBLR for local route inference.
REFERENCES
[1] Daimler. (2016). Mercedes-Benz Intellignet Drive. Accessed:
Aug. 27, 2019. [Online]. Available: https://www.mercedes-benz.com/en/
next/automation/
[2] Daimler. (2016). On the Road in Self-Driving Vehicles. Accessed:
Aug. 27, 2019. [Online]. Available: https://www.daimler.com/case/
autonomous/en/
[3] Hyundai. (2016).Surround Yourself With Innovative Safety, and Supe- rior Driving Dynamics. Accessed: Aug. 27, 2019. [Online]. Avail- able: https://www.hyundaicanada.com/Pages/showroom/PerformanceAnd Safety.aspx?model=genesis
[4] Tesla Autopilot. Accessed: Nov. 28, 2018. [Online]. Available:
https://www.tesla.com/ko_KR/autopilot?redirect=no
[5] W. Brenner and A. Herrmann, ‘‘An overview of technology, benefits and Impact of automated and autonomous driving on the automotive industry,’’
inDigital Marketplaces Unleashed. Berlin, Germany: Springer, 2018, pp. 427–442.
[6] P. Kohli and A. Chadha, ‘‘Enabling pedestrian safety using computer vision techniques: A case study of the 2018 Uber Inc. Self-driving car crash,’’
in Advances in Information and Communication. Cham, Switzerland:
Springer, 2020, pp. 261–279.
[7] S. Thrun, M. Montemerlo, H. Dahlkamp, D. Stavens, A. Aron, J. Diebel, P. Fong, J. Gale, M. Halpenny, G. Hoffmann, and K. Lau, ‘‘Stanley: The robot that won the DARPA grand challenge,’’J. Field Robot., vol. 23, no. 9, pp. 661–692, 2007.
[8] C. Urmson, J. Anhalt, D. Bagnell, C. Baker, R. Bittner, M. N. Clark, J. Dolan, D. Duggins, T. Galatali, C. Geyer, and M. Gittleman,
‘‘Autonomous driving in urban environments: Boss and the urban chal- lenge,’’J. Field Robot., vol. 25, no. 8, pp. 425–466, 2008.
[9] E. Guizzo, ‘‘How Google’s self-driving car works,’’IEEE Spectr., to be published.
[10] J. Dickmann, N. Appenrodt, and C. Brenk, ‘‘Making Bertha,’’ IEEE Spectr., vol. 51, no. 8, pp. 44–49, Dec. 2014.
[11] X. Zhang, H. Gao, M. Guo, G. Li, Y. Liu, and D. Li, ‘‘A study on key technologies of unmanned driving,’’CAAI Trans. Intell. Technol., vol. 1, no. 1, pp. 4–13, Jan. 2016.
[12] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, ‘‘nuScenes: A multimodal dataset for autonomous driving,’’ Mar. 2019,arXiv:1903.11027. [Online]. Available:
https://arxiv.org/abs/1903.11027
[13] K. Jo, J. Kim, D. Kim, C. Jang, and M. Sunwoo, ‘‘Development of autonomous car–part II: A case study on the implementation of an autonomous driving system based on distributed architecture,’’ IEEE Trans. Ind. Electron., vol. 62, no. 8, pp. 5119–5132, Aug. 2015.
[14] K. Jo, J. Kim, D. Kim, C. Jang, and M. Sunwoo, ‘‘Development of autonomous car–part I: Distributed system architecture and development process,’’IEEE Trans. Ind. Electron., vol. 61, no. 12, pp. 7131–7140, Dec. 2014.
[15] K. Jo, M. Lee, J. Kim, and M. Sunwoo, ‘‘Tracking and behavior reasoning of moving vehicles based on roadway geometry constraints,’’IEEE Trans.
Intell. Transp. Syst., vol. 18, no. 2, pp. 460–476, Feb. 2017.
[16] J. Kim, K. Jo, W. Lim, M. Lee, and M. Sunwoo, ‘‘Curvilinear-coordinate- based object and situation assessment for highly automated vehicles,’’
IEEE Trans. Intell. Transp. Syst., vol. 16, no. 3, pp. 1559–1575, Jun. 2015.
[17] T. Bucher, C. Curio, J. Edelbrunner, C. Igel, D. Kastrup, I. Leefken, G. Lorenz, A. Steinhage, and W. von Seelen, ‘‘Image processing and behavior planning for intelligent vehicles,’’IEEE Trans. Ind. Electron., vol. 50, no. 1, pp. 62–75, Feb. 2003.
[18] D. Y. Jeong, S. J. Park, S. H. Han, M. H. Lee, and T. Shibata, ‘‘A study on neural networks for vision-based road following of autonomous vehicle,’’
inProc. IEEE Int. Symp. Ind. Electron., vol. 3, Jun. 2001, pp. 1609–1614.
[19] S. Tsugawa, ‘‘Vision-based vehicles in Japan: Machine vision systems and driving control systems,’’IEEE Trans. Ind. Electron., vol. 41, no. 4, pp. 398–405, Aug. 1994.
[20] Y. Wang, B. M. Nguyen, H. Fujimoto, and Y. Hori, ‘‘Vision-based inte- grated lateral control system for electric vehicles considering multi-rate and measurable uneven time delay issues,’’ inProc. IEEE Int. Symp. Ind.
Electron., May 2013, pp. 1–6.
[21] B.-S. Choi, J.-W. Lee, J.-J. Lee, and K.-T. Park, ‘‘A hierarchical algorithm for indoor mobile robot localization using RFID sensor fusion,’’IEEE Trans. Ind. Electron., vol. 58, no. 6, pp. 2226–2235, Jun. 2011.
[22] Y. Duk-Sun, S. Jae-Heung, K. Min-Seok, pp. Young–Hoon, and K. Jung-Ha, ‘‘The sensor fusion design and performance analysis of unmanned vehicle for the tele-operation system,’’ inProc. IEEE Int. Symp.
Ind. Electron., vol. 3, Jun. 2001, pp. 1425–1430.
[23] J. A. Stover, D. L. Hall, and R. E. Gibson, ‘‘A fuzzy-logic architecture for autonomous multisensor data fusion,’’IEEE Trans. Ind. Electron., vol. 43, no. 3, pp. 403–410, Jun. 1996.
[24] M. Wada, Y. Kang Sup, and H. Hashimoto, ‘‘High accuracy multisensor road vehicle state estimation,’’ inProc. 26th Annu. Conf. IEEE Ind. Elec- tron. Soc., vol. 4, Oct. 2000, pp. 2547–2552.
[25] L. W. Fong, ‘‘Multi-sensor track-to-track fusion using simplified maxi- mum likelihood estimator for maneuvering target tracking,’’ inProc. IEEE Int. Symp. Ind. Electron., Oct. 2007, pp. 36–41.
[26] J. G. Garcia, J. G. Ortega, A. S. Garcia, and S. S. Martinez, ‘‘Robotic software architecture for multisensor fusion system,’’IEEE Trans. Ind.
Electron., vol. 56, no. 3, pp. 766–777, Mar. 2009.
[27] S. M. Han, S. K. Park, J. Jae-Hag, and K. W. Lee, ‘‘Mobile robot navigation by circular path planning algorithm using camera and ultrasonic sensor,’’
inProc. IEEE Int. Symp. Ind. Electron., Jul. 2009, pp. 1749–1754.
[28] T. H. S. Li, Y. T. Su, S. H. Liu, J. J. Hu, and C. C. Chen, ‘‘Dynamic balance control for biped robot walking using sensor fusion, Kalman filter, and fuzzy logic,’’IEEE Trans. Ind. Electron., vol. 59, no. 11, pp. 4394–4408, Nov. 2012.
[29] R. C. Luo and C. C. Lai, ‘‘Enriched indoor map construction based on multisensor fusion approach for intelligent service robot,’’IEEE Trans.
Ind. Electron., vol. 59, no. 8, pp. 3135–3145, Aug. 2012.
[30] R. C. Luo and C. C. Lai, ‘‘Multisensor fusion-based concurrent envi- ronment mapping and moving object detection for intelligent service robotics,’’ IEEE Trans. Ind. Electron., vol. 61, no. 8, pp. 4043–4051, Aug. 2014.
[31] T. Mukai and M. Ishikawa, ‘‘An active sensing method using estimated errors for multisensor fusion systems,’’IEEE Trans. Ind. Electron., vol. 43, no. 3, pp. 380–386, Jun. 1996.
[32] H. Song, W. Choi, and H. Kim, ‘‘Robust vision-based relative-localization approach using an RGB-depth camera and LiDAR sensor fusion,’’IEEE Trans. Ind. Electron., vol. 63, no. 6, pp. 3725–3736, Jun. 2016.
[33] K. Jo and M. Sunwoo, ‘‘Generation of a precise roadway map for autonomous cars,’’IEEE Trans. Intell. Transp. Syst., vol. 15, no. 3, pp. 925–937, Jun. 2014.
[34] D. Lee and W. Chung, ‘‘Discrete-status-based localization for indoor ser- vice robots,’’IEEE Trans. Ind. Electron., vol. 53, no. 5, pp. 1737–1746, Oct. 2006.
[35] P. Bender, J. Ziegler, and C. Stiller, ‘‘Lanelets: Efficient map representation for autonomous driving,’’ inProc. IEEE Intell. Vehicles Symp., Sep. 2014, pp. 420–425.
[36] W. Strijbosch, ‘‘Safe autonomous driving with high-definition maps,’’ATZ Worldw., vol. 120, no. 11, pp. 28–33, Nov. 2018.
[37] H. G. Seif and X. Hu, ‘‘Autonomous driving in the iCity-HD maps as a key challenge of the automotive industry,’’English, vol. 2, no. 2, pp. 159–162, Jun. 2016.
[38] K. Jo, Y. Jo, J. K. Suhr, H. G. Jung, and M. Sunwoo, ‘‘Precise localization of an autonomous car based on probabilistic noise models of road surface marker features using multiple cameras,’’IEEE Trans. Intell. Transp. Syst., vol. 16, no. 6, pp. 3377–3392, Dec. 2015.
[39] K. Chu, J. Kim, K. Jo, and M. Sunwoo, ‘‘Real-time path planning of autonomous vehicles for unstructured road navigation,’’Int. J. Automot.
Technol., vol. 16, no. 4, pp. 653–668, Aug. 2015.
[40] K. Chu, M. Lee, and M. Sunwoo, ‘‘Local path planning for off-road autonomous driving with avoidance of static obstacles,’’IEEE Trans. Intell.
Transp. Syst., vol. 13, no. 4, pp. 1599–1616, Dec. 2012.
[41] J. Kim, K. Jo, K. Chu, and M. Sunwoo, ‘‘Road-model-based and graph- structure-based hierarchical path-planning approach for autonomous vehi- cles,’’Proc. Inst. Mech. Eng., D, J. Automobile Eng., vol. 228, no. 8, pp. 909–928, Jul. 2014.
KICHUN JO (S’10–M’14) received the B.S.
degree in mechanical engineering and the Ph.D.
degree in automotive engineering from Hanyang University, Seoul, South Korea, in 2008 and 2014, respectively. From 2014 to 2015, he was with the ACE Lab, Department of Automotive Engineer- ing, Hanyang University, doing research on system design and implementation of autonomous cars.
From 2015 to 2018, he was with the Valeo Driving Assistance Research, Bobigny, France, working on the highly automated driving. He is currently an Assistant Professor with the Department of Smart Vehicle Engineering, Konkuk University, Seoul. His current research interests include localization and mapping, objects tracking, information fusion, vehicle state estimation, behavior planning, and vehicle motion control for highly automated vehicles.
MINCHUL LEE(S’14) received the B.S. degree in mechanical engineering from Hanyang University, Seoul, South Korea, in 2013, where he is currently pursuing the Ph.D. degree with the Automotive Control and Electronics Laboratory (ACE Lab).
His current research interests include driving envi- ronment assessment, precise positioning and local- ization, information fusion theories, real-time sys- tems for autonomous vehicle systems, and vehicle motion assessment for autonomous cars.
WONTEAK LIM(S’14) received the B.S. degree in mechanical engineering from Hanyang Uni- versity, Seoul, South Korea, in 2013, where he is currently pursuing the Ph.D. degree with the Automotive Control and Electronics Laboratory (ACE Lab). His research interests include trajec- tory planning, vehicle control, real-time systems, intelligent transportation systems, trajectory plan- ning algorithms, a system integration, and soft- ware platform design for autonomous vehicles.
MYOUNGHO SUNWOO (M’81) received the B.S. degree in electrical engineering from Hanyang University, in 1979, the M.S. degree in electrical engineering from The University of Texas at Austin, in 1983, and the Ph.D. degree in system engineering from Oakland University, in 1990.
He joined General Motors Research (GMR) Laboratories, Warren, MI, USA, in 1985, where he has worked in automotive electronics and control for 30 years. During his nine-year tenure at GMR, he worked on the design and development of various electronic control systems for powertrains and chassis. Since 1993, he has been leading research activities as a Professor with the Department of Automotive Engineering, Hanyang University. His current research interests include automotive electronics and controls such as modeling and control of internal combustion engines, design of automotive distributed real-time control systems, intelligent autonomous vehicles, and automotive education programs.