Abstract:Recent advances in video surveillance systems enable a new paradigm for intelligent urban traffic management systems. Since surveillance cameras are usually sparsely located to cover key regions of the road under surveillance, it is a big challenge to perform a complete real-time traffic pattern analysis based on incomplete sparse surveillance information. As a result, existing works mostly focus on predicting traffic volumes with historical records available at a particular location and may not provide a complete picture of real-time traffic patterns. To this end, in this paper, we go beyond existing works and tackle the challenges of traffic flow analysis from three perspectives. First, we train the transition probabilities to capture vehicles' movement patterns. The transition probabilities are trained from third-party vehicle GPS data, and thus can work in the area even if there is no camera. Second, we exploit the Multivariate Normal Distribution model together with the transferred probabilities to estimate the unobserved traffic patterns. Third, we propose an algorithm for real-time traffic inference with surveillance as a complement source of information. Finally, experiments on real-world data show the effectiveness of our approach.

Real-time Traffic Pattern Analysis and Inference with Sparse Video Surveillance Information