Novel Traffic Sensing Using Multi-Camera Car Tracking and Re-Identification (MCCTRI)

Abstract

Traffic sensing devices are the eyes of the Intelligent Transportation Systems (ITS) nowadays. Among all the traffic sensors, the surveillance camera system is one of the most widely deployed system due to the easy installation, valuable data, and the intuitive information format. However, it’s a great pity that these cameras collect data isolated. One camera can only monitor a fixed of view and there is no bridge to share the monitoring information with each other. Tremendous labor work is necessary if the traffic managers try to find the same target in different cameras. Recently, the development of computer vision technology brings light to traffic information extraction based on the multi-camera scenario. Different from the previous single-camera based traffic information estimation, the multi-camera work is much more challenging. Since in the real-world scenarios, different camera views, orientations and lighting conditions make the video features in a huge difference. Moreover, the more rigorous thing is that only the top-one candidate can be used in the traffic information estimation procedure. Thus, how to link each single camera into a multi-camera system and estimate the traffic information from the whole surveillance system becomes the main problem in the research. To address the challenges, four kinds of information are designed to capture and integrate, including vision information, vehicle attributes information, road network graph information and spatial-temporal information. These four kinds of information are summarized and decomposed into four levels of features, including frame-level, clip-level, identity-level and network-level of features. A cutting-edge multi-camera car tracking and Re-ID framework based on temporal-attention model and deep neural networks is improved to capture the frame-level, clip-level and identity-level of features. A Spatial-temporal Camera Graph Inference Model (StCGIM) are designed to integrate the network level of features into the MCCTRI framework. After obtained the multi-camera tracking result, the tracking accuracy levels of different cameras are various from each other. An Adaptative Accuracy Model (AAM) is designed to eliminate and unify errors and prepare the input for the traffic information estimation algorithms. Furthermore, different levels of traffic-related information can be estimated properly. The author evaluated the framework based on five cameras video data on captured on the Interstate 5, including different views, orientations, lighting conditions and color settings in various challenging scenarios. Based on MCCTRI, not only including the traffic information value, such as link average speed, average travel time and volume, but also a more particular data format – the distribution of each parameter can be estimated precisely. All the value information estimation error is less than 8% through the dataset evaluation including five camera views. The KL distance of the estimated distribution and real distribution is less than 3.42. Based on the experiment, the MCCTRI gives the surveillance camera system a brain and more precise and valuable information can be extracted through the method.