PERFORMANCE ON REAL-WORLD DRIVING DATA

The system was 100% accurate at single segment control identification for those identifications that also passed the intersection rules. Table 1 shows the performance results. In the table, ``Raw recognition'' refers to the performance of the classifier alone. Although raw recognition was good, it is not high enough for some applications, such as vehicle safety.

 

Table 1: Filtering the single segment control identifications using the multi-segment intersection rules yields 100% accuracy.

	Instances	Correct	Percent
Raw Recognition	100	90	90.0
Intersection Threshold	59	56	94.9
Legal Ground Truth	57	54	94.7
Legal Test Results	50	50	100.0

 

Some intersections only had one entering segment that passed the minimal sampling criteria. Since these intersections trivially satisfy one of the intersection rules, they were excluded. ``Intersection Threshold'' refers to the requirement that at least two segments be available to include all the segments of an intersection.

In rare cases, the ground truth for some segments violated the intersection rules. ``Legal Ground Truth'' refers to system performance after the system has rejected these segments.

Finally, ``Legal Test Results'' refers to the system's performance after it had rejected results for test data that violated the intersection rules. A sample legal intersection is all four entering segments classified as ``stoplight.'' A sample illegal intersection is three segments classified ``stoplight'' and one classified ``stop sign.''