Improved kernel density background estimation with diversity sampling and neighbor information for traffic monitoring

Abstract

Dynamic background extraction is one of the key tasks in moving object detection in static camera surveillance for traffic monitoring. The kernel density estimation is for nonparametric multi-mode background modeling with the advantage of dealing with waving tree leaves, high frequency and repeated motion etc. To its computation expensive with repeated similar samples, an improved nonparametric background model using novel diversity-sampling mechanism is proposed to extract the important and diverse samples. In the learning phase, several samples having more popular and various intensity values are extracted from the original sample set. Different weights are given to the distinct samples according to the related intensities. In the evaluation phase, the underlying probability density function is estimated using the weighted diversity samples in kernel density estimation. The diverse sample-set makes the evaluation computation inexpensive and efficient. The effectiveness of the proposed method is demonstrated in the traffic monitoring application. © 2012 Springer-Verlag Berlin Heidelberg.