Wireless Sensing and Analytics
7.4 Experimental Results
7.4.1.5 Long-Term Test in House #2
Furthermore, the long-term behavior of the presented event detector in TRIMS is inves- tigated in House #2 through a test that lasts for 2 weeks. During the long-term test, resident activities are more often than that in House #1, and thus the indoor environ- ment changes every day, which might jeopardize the presented event detector trained in day 1. Every day during the long-term test, the tester performed the same operational test as inSection 7.4.1.4to evaluate the detection performance of the presented event detector. The system outputs along the time are plotted inFigure 7.9,where they-axis is the system output with “unknown,” “Allclosed,” “Front,” and “Back” representing the occurrence of untrained events, e1 “all doors are closed,” e2 “front door is opened,” and e3 “back door is opened,” respectively.
As shown inFigures 7.9(a), 7.9(f),and7.9(k),the presented event detector is good at detecting the trained events with no false alarm during the same day when the system is trained. However, after 1 week or even 2 weeks, with the original training database built on day 1, the presented system fails to detect the trained events and has a high false alarm rate on e2, as shown in Figures 7.9(b), 7.9(d), 7.9(g), 7.9(i), 7.9(l),and7.9(n).
7.4 Experimental Results 163
50 100 150
Unknown Allclosed Front Back
Time (s) (a)
50 100 150
Unknown Allclosed Front Back
Time (s) (b)
50 100 150
Unknown Allclosed Front Back
Time (s) (c)
50 100 150
Unknown Allclosed Front Back
Time (s) (d)
50 100 150
Unknown Allclosed Front Back
Time (s) (e)
20 40 60 80
Unknown Allclosed Front Back
Time (s) (f)
20 40 60 80
Unknown Allclosed Front Back
Time (s) (g)
20 40 60 80
Unknown Allclosed Front Back
Time (s) (h)
20 40 60 80
Unknown Allclosed Front Back
Time (s) (i)
20 40 60 80
Unknown Allclosed Front Back
Time (s) (j)
20 40 60 80
Unknown Allclosed Front Back
Time (s) (k)
20 40 60 80
Unknown Allclosed Front Back
Time (s)
(l)
20 40 60 80
Unknown Allclosed Front Back
Time (s) (m)
10 20 30 40 50 60 Unknown
Allclosed Front Back
Time (s) (n)
10 20 30 40 50 60 Unknown
Allclosed Front Back
Time (s) (o)
Figure 7.9 Monitoring results of the event detector for long-term tests in House #2. (a) e1 test on day 1. (b) e1 test on day 7 w/o e1 update. (c) e1 test on day 7 w/e1 update. (d) e1 test on day 14 w/o e1 update. (e) e1 test on day 14 w/e1 update. (f) e2 test on day 1 (open around the 20th and the 62th s and close around the 34th and the 74th s). (g) e2 test on day 7 w/o e1 update (open around the 20th and the 60th s and close around the 30th and the 70th s). (h) e2 test on day 7 w/e1 update. (i) e2 test on day 14 w/o e1 update (open around the 18th and the 58th s and close around the 30th and the 70th s). (j) e2 test on day 14 w/e1 update. (k) e3 test on day 1 (open around the 18th and the 52th s and close around the 24th and the 62th s). (l) e3 test on day 7 w/o e1 update (open around the 46th and the 72th s and close around the 52th and the 80th s).
(m) e3 test on day 7 w/e1 update. (n) e3 test on day 14 w/o e1 update (open around the 12th and the 40th s and close around the 20th and the 48th s). (o) e3 test on day 14 w/e1 update.
For example, as depicted in Figure 7.9(b), the system keeps reporting “front door is opened,” when the ground truth of the indoor state being e1 “all doors are closed.” With uncontrolled resident activities, the indoor environment changes result in a different multipath profile not only for e1 but also for e2 and e3. With the help of the auto-update of e1, the presented event detector is able to detect the trained events e2 and e3 during the 2-week experiment with no false alarm. The results are as shown inFigures 7.9(c), 7.9(e), 7.9(h), 7.9(j), 7.9(m),and7.9(o).
As demonstrated by examples inFigure 7.9,with an automatic and periodic update of the training data for e1, TRIMS can maintain its accuracy in differentiating between and recognizing trained events in a single-family house with normal resident activities during 2 weeks. The monitoring results of TRIMS in the 14-day experiment are com- pared with the history log provided by a commercial home security system. In general, the presented event detector captures the incidents of e2 and e3, i.e., opening the front or the back door from the outside of the house, with an accuracy being 95.45% while a single misdetection happens on day 13.
7.4.2 TRIMS: Motion Detector
The performance of the motion detector is tested in House #1 with the TX and RX devices located at positions marked by “TX_2” and “RX_2” in Figure 7.5(a). The parameter αdefined in (7.25) is set to be either 0.8 or 0.2, whileW = 30 indicates a 1-s window of continuously collected CSI as defined in (7.24). For theDmot ion(t0) in (7.26), we apply a time-diversity smoothing with only one-level majority vote whose W =45 andO =15, to eliminate any possible false alarms due to burst noise or error in the CSI estimation.
During the training phase, the presented motion detector learns the thresholdγmot ion
based on the data from 1-min monitoring data collected under e1 and someone walking in and around the center of the house.
Given a zero false alarm, the detection rates of motion at different locations are listed in Table 7.3. The presented motion detector is intelligent in that it learns and adapts its sensitivity automatically based on the characteristics of the radio propagation environment where it is deployed, through the training phase. The change that motion Table 7.3 Detection rate for motion at different locations under 0 false alarm rate.
Walking location Detection rateα=0.8(%) Detection rateα=0.2(%)
Postman 0 0
Foyer 59.18 12.24
Laundry room 100 81.25
Alice’s room 3.92 0
Study room 1.92 0
Center of the house 83.33 75.93
Kitchen 48 36
Living room 30.91 0
Restroom 0 0