Direction-Oriented Topic Modeling with Applications in Traffic Scene Analysis

Unlike text analysis for which topic models are historically developed, traffic video analysis is dealing with much simpler topics, made of restricted motion patterns. In this paper, we propose a dual-layer direction-oriented framework for more efficient traffic motion patterns description based on topic models through considering the simplicity of traffic topics. The aforesaid framework compels the involved topic models to learn the foreknown visually meaningful motion patterns that exist in traffic scenes, as developed theoretically in this paper. Experimental results produced by common datasets show that the proposed method provides more intuitive topics for traffic flow description. Based on experimental results, our framework outperforms other topic-model based methods by 4% to more than 11% in detecting abnormal events, in terms of the area under the Receiver Operating Characteristic curve. In addition to that, in a scene analysis evaluation at intersections equipped with traffic signals, our method reaches 4% higher traffic phase detection accuracy, compared to conventional topic models.