Asynchronous Bayesian algorithm for object classification: Application to pedestrian detection in urban areas

This paper considers the problem of the classification of objects observed by vehicle embedded sensors. We propose a general architecture and an algorithm to perform multisensor fusion for the classification purpose. The proposed solution has to be robust and flexible. The robustness is essential because this system is for safety applications. The flexibility is ensured by a modular architecture alongside with a feature-level data fusion algorithm. This algorithm is based on the Bayesian formalism. Therefore, the classification of a detected object becomes the computation of a recognition probability given a set of features. This formalism allows an asynchronous data fusion, that is to say each sensor information is taken into account as soon as it is available. The proposed approach is applied to combine information from a laser scanner and video camera in order to detect pedestrians moving in the equipped vehicle path. The experimental results present pedestrians detected in urban areas. These results confirm the effectiveness of the proposed approach.