Traffic monitoring using short-long term background memory

Abstract

Background subtraction is an efficient technique in vision-based traffic monitoring, it segments the moving vehicle from the video sequences by comparing the incoming frame to the model of background scene. The presented work is a simple approach of adaptive background modeling in which the short term memory (STM) and long term memory (LTM) are introduced to construct the whole background memory. The color cue is used to build the model of pixel, u* and v* chrominancy components are carefully selected from modified L*u*v* color space, they are perceptually uniform such that color difference could be measured properly. Furthermore, object shadows are suppressed because the luminancy effects are removed. A simple prototype cell is defined to characterize the background scene by its 'circular influence field'. The match of prototype cell is measured by the Euclidean distance between the incoming pixel and prototype cell. The most recent prototype cells are stored in STM, they adapt quickly for the variations of background scene, but false detections easily occurs when the background has the high frequency variations. In LTM, prototype cells store the stable representation of the background scene, which are able to reduce the computation of STM updating. The adaptive learning procedure is carried out in both STM and LTM, it is able to deal with the scene changes. This background model is evaluated by the traffic video stream, experimental results show that the proposed approach is feasible for the traffic monitoring.