Vision-Based Occlusion Handling and Vehicle Classification for Traffic Surveillance Systems

Abstract

Due to the factors such as visual occlusion, illumination change and pose variation, it is a challenging task to develop effective and efficient models for vehicle detection and classification in surveillance videos. Although plenty of existing related models have been proposed, many issues still need to be resolved. Typically, vehicle detection and classification methods should be vulnerable in complex environments. Moreover, in spite of many thoughtful attempts on adaptive appearance models to solve the occlusion problem, the corresponding approaches often suffer from high computational costs. This paper aims to address the above mentioned issues. By analyzing closures and convex hulls of vehicles, we propose a simple but effective recursive algorithm to segment vehicles involved in multiple-vehicle occlusions. Specifically, a deep convolutional neural network (CNN) model is constructed to capture high level features of images for classifying vehicles. Furthermore, a new pre-training strategy based on the sparse coding and auto-encoder is developed to pre-train CNNs. After pre-training, the proposed deep model yields a high performance with a limited labeled training samples.