PNN Based Motion Detection with Adaptive Learning Rate

This paper proposed a new motion detection algorithm based on neural network (NN). Video background was modeled by combing probabilistic neural network (PNN) and winner take all (WTA) network, which is called adaptive background PNN (ABPNN). Every pixel in a video frame was classified to be foreground or background by conditional probability of being a background. Foreground was further classified into motion region and shadows by shadow detection. Background probability was estimated by a Parzen estimator in HSV feature space. Both Parzen estimator and network weights were updated online according to classification results, and weight learning rate was adapted according to ratio of motion regions. Experimental results on benchmark videos show that the proposed algorithm can detect motion more precisely than other NN based method, and it can also adapt to sudden background changes more quickly.