ISSN 2250-0987 Tapan kumar das et al, UNIASCIT, Vol 1 (3), 2011, 107-113 107 A Survey of Techniques for Background Subtraction and Traffic Analysis on Surveillance Video U.Chandrasekhar 1 ,Tapan kumar Das 1 1 School of Information Technology & Engineering , VIT University, Vellore,India Abstract—Identifying stationary background from moving objects in a video is a critical task in many video processing and computer-vision applications. A basic approach is to perform background subtraction to isolate moving objects from a significant background model. The main challenge in this task is to devise an algorithm to work like a human would, under different scenarios and context. Technically, robustness of such algorithm will be tested by varying illumination, identifying non-stationary background objects like swinging leaves, rain, snow and shadow of moving objects. Also context wise, the algorithm should react quickly to changes in background like parked vehicles and background objects brought in later into the scene. Since we are interested in analyzing traffic frequency in a surveillance video, other challenges include identifying non vehicular moving objects like pedestrians etc. In this paper we compare various simple background verification techniques like frame differencing, Gaussian methods to few complex techniques like probabilistic models. Though complex techniques produce superior and context specific results, our experiments show that, for simple tasks like traffic analysis, simple techniques like adaptive median filtering produce good accuracy with low processing time. Video surveillance of traffic happens from a stationary camera having constant field of view. Generally the resolution and frame rate of these videos are low for high end algorithms to produce good results. Also traffic analysis is not required on live video stream. Except for spontaneous high security alerts, processing can be done offline by buffering the stream data. Hence, the paper concentrates on methods which are robust enough to handle noise, changing climatic conditions and issues with segmentation of moving objects. Keywords- Background subtraction, Traffic video analysis I. INTRODUCTION Today there are an ever growing number of cameras being used for scene analysis. Many of these are applied to traffic monitoring because it is a low cost and passive method for data collection. Research in several fields of traffic applications has resulted in the wealth of video processing and analysis methods [1]. Automatically detecting and tracking vehicles in video surveillance data is a challenging problem in computer vision with important practical applications, such as traffic analysis and security. Manually reviewing the large amount of data they generate is often impractical. Thus, algorithms for analyzing video which require little or no human input are an attractive solution and have been an area of active research for over a decade. In addition to correctness in detecting and tracking vehicles, the computational complexity of a tracking system is important. For many applications, real-time or near real-time tracking capability is desired. There are various algorithms for object recognition in video. Sample areas include object tracking, traffic monitoring and analysis, Human tracking and gesture recognition etc. Almost all the projects involve identifying background from foreground objects. Also this is the most challenging part. In cases like object tracking, background is continuously changing and in cases like video surveillance background is constant and foreground objects are in motion. This work describes a single stationary camera feed on which the processing techniques are carried out. A comparison of the techniques and their accuracies, merits and demerits are also given. Fig.1. Flow diagram of a general background estimation process