Moving objects detection and credal boosting based recognition in urban environments

In this paper, we propose a stereo-vision based system for moving objects detection and pedestrian recognition using Residual Image Motion Flow (RIMF) and Credal Boosting algorithm. First, the RIMF is computed to obtain residual motion likelihood for each possible image pixel. Next, depth-based hierarchical Mean Shift clustering algorithm is applied to generate bounding boxes around the moving objects. Once we obtained the bounding boxes of moving objects, we propose, for the recognition step, to focus on pedestrian. For this purpose we use a Credal Boosting classification algorithm which make a decision based on the degrees of credibility provided by weak learners. Compared to classical classifiers, Credal Boosting classifier has a lower false negative rate and also provide more information in the form of class and belief functions on them. Several experimental results show that our system can detect dynamic objects in complex urban traffic scenes.