Algorithm Design And System Optimization Of Vehicle-mounted Thermal Imaging Pedestrian Detection

As one of the key technologies of the Automotive Driver Assistance System,vehicle-mounted thermal imaging pedestrian detection has important academic significance and practical application value.Although a variety of vehicle-mounted thermal imaging pedestrian detection algorithms have been proposed under the framework of machine learning,achieving real-time and robust performance is still a challenging issue.On the one hand,pedestrian detection needs to handle dynamic and complicated traffic scenes,and the intra-class diversity of pedestrian are large.On the other hand,although thermal imaging highlights pedestrian targets in images,useful information such as color and pedestrian details are lost heavily.In this paper,focus on the algorithm design and system optimization of vehicle-mounted thermal imaging pedestrian detection.The main work includes:1)Probability map based regions of interest(RoIs)extraction method for vehicle-mounted thermal imaging pedestrian detection is proposed,which includes image preprocessing,probability map calculation and RoIs generation.For image preprocessing,using convex-concave gray value mapping curve maps image pixels to enhance image contrast.For probability map calculation,two kinds of probability map are proposed,the final probability map is obtained by fusion of the two.For RoIs generation,a search algorithm which uses pedestrian prior knowledge and statistical distribution as the heuristic information or constraints is proposed to speed up the process and improve the accuracy of RoIs positioning.2)Multi-feature fusion based classifier for vehicle-mounted thermal imaging pedestrian detection is proposed.It improves the performance of the classifier by optimizing feature representation ability,and the feature representation ability is optimized by multiple features fusion.The multi-scale feature fusion template is proposed to indicate different features to be extracted in different scales and regions of images.Two-stage feature selection method is proposed to obtain multi-scale feature fusion template.3)The detection process of the vehicle-mounted thermal imaging pedestrian detection system is optimized to integrate the pedestrian detection algorithm efficiently.A probability map based pedestrian tracker is proposed to improve the tracking performance and realize data multiplexing.