Performance Matrix of Tree Detection Module (Distance)
6.5 Algorithm: Fusing Water Cue
Water bodies may appear in different kind of appearances thus is not expected that no single features can be used to detect the entire water body. In this thesis, multiple-feature approach is taken to enhance the detection of water bodies with multiple cues. Previous sections discusses features that can be used to target specific water attribute in the while minimizing the false detection.
Figure 6.9 shows the framework of water detection module in this thesis.
The input images are polarized stereo camera image with difference polarization angle between the image pair. Stereo processing is performed on the image pair to produce disparity data. Note the detected ground plane is also fed into this module to reduce the false detection of the water bodies. Any detection of water region above the ground plane will be filtered as water bodies presence will only occurs on the ground plane.
There are four sub-modules which target different type of water attributes. The sub-modules are sky-reflection detection based on color and
brightness information, low-texture region detection, invalid disparity detection and object reflection detection. Once the sub-modules detect the water body region of interest, the region is refined by post-processing. The following sections discuss in details the operation of each sub-module.
Figure 6.9: Framework for water bodies detection using multiple features.
6.5.1 Sky Reflection Detection Module
The RGB images selected from our archive for processing were converted to hue, saturation, and value (HSV) color space. There are several factors that contribute to the surface color of water bodies. Among them
Input Polarized - Stereo Image
Left Polarized - Stereo Image
Right Polarized - Stereo Image
Disparity Image
Sky Reflection Detection
Low- Texture Detection
Invalid Disparity Pixel Detection
Water Region
Estimated Ground Plane
Object Reflection Detection
Update Estimated Ground Plane Post- Processing - Region Growing and Connected Component
the color of the sky reflecting on the water, the color of background material casting a shadow on the water, and whether or not the water is moving. As these factors have great variation, it is difficult to predict the hue of water (Rankin, Matthies, & Huertas, 2004). Note that the reflection of the sky in water has low saturation values and high brightness values.
In this sub-module, we follow closely the work by Rankin et al (2004).
The following are the rules imposed in this work:
6.12
where, is saturation, is brightness and is hue. Rule 1 and rule 2 targets the sky reflection characteristics where it has low saturation and high brightness value. Rule 3 is lower brightness thresholds are applied only if the sky is detected in the imagery. Rule 4 is the only one that uses hue. It targets deep bodies of water, which tend to have a blue hue.
Note that in his approach, the sky detection is done the top ten rows of the image. Rule 3 and rule 4 are only activated if the sky is detected. However in this thesis, the sky detection is not performed as the search for water bodies in the scene is constrained to the region below the horizon. This is due to the
nature of water body to be on the ground plane rather than areas of higher elevation. Thus, false detection such as sky as water bodies can be eliminated.
6.5.2 Low-Texture Detection Module
Texture is an important approach for region description. Although no formal definition of texture exists, intuitively this descriptor provides measure of properties such as smoothness, coarseness and regularity. We use one of the simplest approaches to describe texture which is variance. Variance can be used in texture description as it is a measure of intensity contrast and is give as
6.13
where is the number of samples and is the intensity of the pixel. It can be used to establish descriptors of relative smoothness using the following
6.14
A 3x3 intensity variance filter is passed over grayscale image. At each pixel, the window variance and relative smoothness is calculated. Then the water body region with relative smoothness less that threshold set will be selected.Table 6.2 shows the values of the relative smoothness of water bodies
objects where the relative smoothness is rather coarser. Based on the average values, we formulate the threshold that will be used to detect water bodies cue and differentiate it from other objects.
Objects Average R(normalized) Threshold Set for R
Water Bodies ~0.001 R < 0.001
Other Objects ~0.079 R > 0.01
Table 6.2: Average relative smoothness for 20 samples for water bodies and other objects in rainforest scene.
6.5.3 Object Reflection and Invalid Disparity Pixel Detection Module.
In this module, the partial polarization feature is used to remove the specular reflection in one of the stereo image. In Section 6.4.1, we have shown that water bodies will reflect more light compared to other objects. Using these characteristics, we will be able to distinguish water body from other regions.
In this thesis, partial polarization of is used since the magnitude of partial polarization for water is highest at this angle. When different polarization angle is applied to the stereo camera, the image pair appears to be different. Therefore, the stereo correspondence cannot be found. Consequently, there will be no disparity data for the water region.