A Study Of Background Subtraction In Urban Traffic

Object detection in urban traffic is a common application in intelligent surveillance systems. High accuracy in object detection and image segmentation is very important to upper layer intelligent applications such as violation recognition, parking recognition and traffic statistics. This paper focuses on object recognition theory in urban traffic. Firstly, this paper concludes commonly used object recognition methods. Several classic background modeling methods are presented with their principles and advantages & disadvantages. Several main problems in background modeling such as "Moved Objects”, “Time of Day”, "Bootstrapping" and "Foreground Aperture" are concluded and analyzed in detail. In order to avoid these problems in urban traffic, this paper proposes a random update based codebook modeling method which is more accurate and robust in urban traffic object detection system. The main contributions of this paper includes:(1)Analysis of classic background modeling methods including Average Background Modeling, Gaussian Mixture Background Modeling, Kernel Density Method, Vi BE Method and Codebook Method. Advantages & disadvantages as well as suitable scene are presented for each method. The main characteristics of urban traffic are: no "clean" frames available for background model’s training, gradual illumination change and temporary halting of foreground objects. Improvements have been made to solve the problems above.(2)A random update based Codebook method is proposed, the main improvements are: the introducing of random strategy in the real time updating of each codeword, the random strategy makes each pixel value be chosen randomly based on its reliability. as a result of random updating strategy, the former detection miss caused by too high a updating rate and bad adaption ability caused by too low a updating rate are avoided, and the background model becomes more reliable and robust. On the basis of statistics analysis, regular pattern of background codeword is generated, statistics and sorting method are used to filter out the reliable background codeword. The "bootstrapping" problem is solved perfectly by this statistics method. Space consistency information is used to speed up the updating of the background region as the background region is more likely to be of high space consistency, it makes the background model more stable and temporary parking vehicles won’t be incorrectly undetected.(3)Several experiments have been made on both standard datasets and real urban traffic datasets, ROC is generated to analyze the performance of new method at the same time.It is proved by experiments that the random update based Codebook modeling method has the ability to build the background reference image accurately despite the impact of moving foreground objects along the training period, and the background model is adaptive to the gradual illumination change. This method has good performance in urban traffic scene.