Beyond bag-of-words: An improved Sparse Topical Coding for learning motion patterns in traffic scenes

Analyzing motion patterns in traffic videos can directly generate some high-level descriptions of the video content which can be further employed in rule mining and abnormal event detection. The most recent and successful unsupervised approaches for complex traffic scene analysis are based on topic models. However, most existing topic models share some key characteristics which could limit their utility. In this paper, based on extracted optical flow features from video clips, we employ Sparse Topical Coding (STC) framework to automatically discover typical motion patterns in traffic scenes. For this purpose, we improve the STC to overcome one of the drawbacks of topic models with the aim of learning the semantic traffic motion patterns. We go beyond the usual word-document paradigm in topic models by taking into account the order of optical flow words during learning. Experimental results show that our proposed method can learn better motion patterns to analyse the traffic video scenes.