Fast and Robust Background Updating for Real-time Traffic Surveillance and Monitoring

Abstract

Background updating is an important aspect of dynamic scene analysis. Three critical problems: sudden camera perturbation, sudden or gradual illumination change and the sleeping person problem, which arise frequently in realworld surveillance and monitoring systems, are addressed in the proposed scheme. The paper presents a multi-color model where multiple color clusters are used to represent the background at each pixel location. In the proposed background updating scheme, the updates to the mean and variance of each color cluster at each pixel location incorporate the most recently observed color values. Each cluster is assigned a weight which measures the time duration and temporal recurrence frequency of the cluster. The sleeping person problem is tackled by virtue of the observation that at a given pixel location, the time durations and recurrence frequencies of the color clusters representing temporarily static objects are smaller compared to those of color clusters representing the true background colors when measured over a sufficiently long temporal history. The camera perturbation problem is solved using a fast camera motion detection algorithm that allows the current background image to be registered with the background model maintained in memory. Sudden illumination changes are handled by using an adaptive histogram template whereas gradual illumination changes are automatically resolved with the adaptive background model. The background updating scheme is shown to be robust even when the scene is very busy and also computationally efficient, making it suitable for realtime traffic surveillance and monitoring systems. Experimental results on real traffic monitoring and surveillance videos are presented.