Design and implementation of a stereo vision system for vehicle detection and classification

Abstract

This master thesis presents the implementation of an automatic door sensor. This work is based on stereoscopy which provides the information on depth thanks to a stereo camera. It was divided into several steps. Each step of the process was conscientiously realized and optimized to deliver the best possible outcome. The first steps concern the intrinsic and extrinsic calibrations of the cameras. Once both calibrations have been performed, the disparity map is computed. The disparity is the difference in image location of an object between the right and the left images. Then a background subtraction is performed to keep only the relevant parts of the scene. The chosen method is called ViBe, developed by M. Van Droogenbroeck, O. Barnich and O. Paquot. After that, the foreground of the scene is projected in 3D. The 3D point cloud is directly downsampled using a VoxelGrid filter in order to have a constant density. The features in the scene are differentiated and then tracked from one frame to the next. Their positions and trajectories are computed in order to make decisions concerning the opening of the door. The sensor does not only open or close a door. Two people counters and a classification algorithm were implemented. The first counter counts the number of people entering and exiting the door and the second one counts the number of people that passes in front of the door whether they cross it or not. These counters can be used to generate data for marketing strategies. The classification, between these three classes : {human, car, truck}, aims at reducing the heat loss in warehouses by opening the door at an appropriate height but also at collecting data. The 3D vision technology is promising for new generation sensors. The sensor presented in this work showed good results, which demonstrates the potential of this technology.