Fast and Robust Object Detection and Tracking for Lightweight Visual Surveillance Systems

In this paper, we present a novel method for moving object detection using background subtraction and tracking for lightweight visual surveillance systems. The proposed moving object detection method using background subtraction is based on a “Gaussianity test” and a shading model. The method has been shown to be robust to dynamic scenes such as sudden illumination changes and presence of relatively small scene clutter motion (e.g. waving tree branches and leaves). A fast implementation of the method using an extension of integral images is presented. A fast implementation of object tracking using a particle filtering framework is also presented using a probability density function of the visual features of the target object. A weighting scheme based on non-uniform kernel filtering of the probability density function is described for better tracking performance. We also show in our proposed background subtraction method that the construction of the object tracking appearance model has speed improvements of orders of magnitudes compared to traditional methods such as convolution.