Bounding Box Accuracy in Pedestrian Detection for Intelligent Transportation Systems

This paper describes a stereo-vision-based pedestrian detection system for intelligent transportation systems. The basic components of pedestrians are first located in the image and then combined with a SVM-based classifier. Generic obstacles are located using a subtractive clustering attention mechanism based on stereo vision. A by-components learning approach is proposed and different feature extraction methods are tested in order to better deal with pedestrian variability and justify what features are better to be learnt for pedestrian detection. Candidate selection mechanisms usually yield pedestrians with inaccurate bounding boxes. Then a decrease in detection rate takes place if the SVM classifier is trained only with well-fitted pedestrians. Using several off-line databases containing thousands of pedestrians samples the effect of bounding box accuracy is studied. A multi-candidate generation mechanism is also developed in order to enhance the single frame performance, decreasing the number of false positives due to inaccurate bounding boxes