Spatiotemporal saliency detection in traffic surveillance
Wei Li, Dhoni Putra Setiawan, Hua-An Zhao · 2017
Moving vehicle segmentation in traffic videos is a challenging work because of complex background and variety objects. In this paper, we focus on detecting vehicles that are running through crossroads using the up-to-date spatiotemporal saliency model. The current saliency detection methods aim at detecting the most salient objects, novel but stationary target will be easily classified as foreground, which is a misclassification in moving object detection. We propose a new set of appearance and motion feature and an improved optimization model to solve this problem. During the procedure of saliency map calculation, motion information is treated as a more important role compared to spatial feature. Therefore, moving objects can be segmented easier. Some experimental results showed, compared to a current method, our approach could segment moving vehicle more precisely.