Efficient Analysis of Traffic Intersection Scenes by Employing Traffic Phase Information

Analyzing motion patterns in traffic videos can be employed directly to generate high-level descriptions of their content. For traffic videos captured from intersections, usually, we can easily provide additional information about traffic phases. Such information can be obtained directly from the traffic lights or through traffic lights controllers. In this paper, we focus on incorporating additional information to analyze the traffic videos more efficiently. Using side information on traffic phases, the semantic of motion patterns from traffic intersection scenes can be learned more effectively. The learning is performed based on optical flow features extracted from training video clips, and applying them to supervised topic models such as MedLDA and MedSTC. Based on such models, any video clip can be represented based on the learned patterns. Such representations can be further exploited in scene analysis, rule mining, abnormal event detection, etc. Our experiments show that employing side information in intersection video analysis leads to improvement in discovering scene pattern. Moreover, supervised topic models achieve about 4% improvement in abnormal event detection, compared to the unsupervised ones, in terms of area under ROC.