Feature-based Multisensor Fusion Using Bayes Rule for Pedestrian Classification in a Dynamic Environment

This paper describes how multisensor data fusion increases reliability of pedestrian classification while using a Bayesian approach. The proposed approach fuses information provided by a laser range scanner and a monocular grey-level camera. Fusion is applied at feature level by using sets of related features and possibly correlation sensor observations. The clue is to combine in a probabilistic framework, using Bayes rule, the detecting capabilities of sensors used in order to be able to detect and locate the presence of pedestrians along the vehicle trajectory. To do that, three main processes are performed, namely, sensor data pre-processing, tracking and classification. This work emphasizes the idea of redundancy due to the different nature of the information provided for addressing classification task by different sensors. The pedestrian classification benefits from pattern classification, a priori known static and dynamic restriction of the pedestrians under consideration. Contributions brought are : estimation of likelihoods, P(feature|class), which is defined as the likelihood a detected object belongs to an object class (pedestrian/non pedestrian) according to observed feature ; gain in reliability of pedestrian classification brought by likelihoods combination based on Bayes rule ; and simplicity to integrate of all past knowledge and also using Bayes’theorem. To estimate the benefits, we used simulation based on the Bayesian statistic framework. We demonstrate with the help of ROC curves (good detection rate versus false positive rate) the gain obtained by multisensor classifier compare to single sensor classifier approaches. Finally, experiments using real data processed in an off-line process confirm the expected results.