Section: New Results
Deformable Parts Model based approach for on-road object detection and classification
Participants : Wei-Lin Ku, Evangeline Pollard, Anne Verroust-Blondet.
An important perception problem for driver assistance is the detection of the road obstacles and the recognition of their type (cars, cycles, pedestrians). This year, we tackled the on-road objects detection problem by testing and improving vision-based methods. We proposed and compared several DPM based strategies for on-road object detection and classification, laying emphasis on the problem of detecting smaller/occluded cars and pedestrians. A hybrid approach combining detection from small/large models trained with different clustering method has been introduced to boost the detection performance in both Average Precision and Maximum Recall in every difficulty level. Finally, a geometry reasoning based filtering has been employed to eliminate false alarms while preserving a great deal amount of true positives. Experimental results showed the improvement both in hybrid and geometry reasoning approaches. Most of this work has been done during the internship of Wei-Lin Ku.