AbstractsComputer Science

Vehicle detection and tracking in road traffic surveillance is a classical task in computer vision and a critical component in a modern traffic monitoring system. However, most traditional systems are rendered ineffective in the face of three fundamental challenges inherent in visual object detection and tracking systems: occlusion, appearance variation and tracking failure. Compared to other object classes, the vehicle class exhibits more severe occlusions due to the low-angle low-height nature of road traffic surveillance cameras and the huge population of vehicle instances, and more appearance variation resulted from the high diversity of that population. The situation of tracking is no better owing to the complex environments in traffic scenes and the occlusion problem. We explore two kinds of features, namely a vehicle-level feature - windshield, and a low-level feature - interest point. Both of the two features address the three challenges to some degree. Windshields observe less occlusions and a more uniform appearance than vehicles, and tracking is consequently easier; Interest point is a very stable and distinctive local image feature which is occlusion robust, and detection by grouping makes appearance variation irrelevant. Besides working independently, the heterogeneity of the two features also makes their collaboration beneficial. The semantic information of windshield provides guidance to the grouping of interest point, while the stability of interest point is an indispensable compensation to the failure modes of windshield. We develop systems based on the individual features as well as the integration of features. We evaluate our systems on real world video sequences, and the results demonstrate the advantages of the proposed algorithms and their superior performances over the state-of-the-art methods.