Adaptive Bayesian Combination of Features from Laser Scanner and Camera for Pedestrian Detection

This paper describes how multisensor data fusion increases reliability of pedestrian detection while using a Bayesian combination of features. The clue is to combine in a probabilistic framework, the detecting capabilities of sensors for identifying pedestrians located along the vehicle trajectory. The work emphasizes the idea of redundancy due to the different nature of the information provided by laser scanner (a priori static outlines and dynamic restriction of the walking pedestrian) and camera (pattern classification) for addressing pedestrian detection. Bayesian classification consists of computing for each detected object, probability of being a pedestrian and compare it to a predefined threshold. Contributions brought are estimation of sensor models, p(feature|object class), based on heuristics and training processes; an original way to take into account scale variation in vision model in order to improve classification of far or small objects; and integration of past knowledge when processing sequences to enhance classification accuracy. Performance of vision-, laserand combined features-based classifier is analyzed by means of Receiver Operating Characteristics (ROCs). Features combination provides an optimized system. Experimental results using real data (performed in an off-line process) suggest a Bayesian combination of features as an essential clue to enhance performance of pedestrian detection system.