Feature-based Multisensor Fusion Using Bayes Formula for Pedestrian Classification in Outdoor Environments

Improvements on pedestrian classification reliability applying a Bayesian approach to multisensor data fusion is described in this paper. The proposed approach fuses information provided by a laser scanner and a monocular gray-level camera. The key is to combine in a probabilistic framework, the detecting capabilities of these sensors to classify pedestrians located along the vehicle trajectory. The approach comprises three processes: sensor data processing, tracking and classification. This work emphasizes the idea of redundancy and complementarity due to the different nature of the information provided by the laser scanner (a priori static outline and dynamic constraints of the pedestrian motion) and camera (patterns) to address pedestrian classification. Two contributions are presented: 1) estimation of likelihoods, ^(feature class), which is defined as the likelihood that a detected object belongs to a class (pedestrian or non-pedestrian) according to an observed feature; 2) likelihood combinations as well as past knowledge integration using Bayes formula. The performance of vision, laser and combined feature-based classifier through the application of a Receiver Operating Characteristics (ROCs) analysis is included. It was found that the combination of features results in an optimized system. Experimental results using real data (performed off-line) suggest that a Bayesian combination of features is an essential factor to enhance performance of pedestrian detection systems.