An Object Tracking Scheme Based on Local Density

We propose a method for tracking an object from a video sequence of moving background through the use of the proximate distribution densities of the local regions. The discriminating features of the object are extracted from a small neighborhood of the local region containing the tracked object. The object’s location probability is estimated in a Bayesian framework with the prior being the approximated probabilities in the previous frame. The proposed method is both practical and general since a great many of video scenes are included in this category. For the case of less-potent features, however, additional information from such as the motion is further integrated to help improving the estimation of location probabilities of the object. The nonstatistical location of an object is then derived through thresholding and shape adjustment, as well as being verified by the prior density of the object. The method is effective and robust to occlusion, illumination change, shape change and partial appearance change of the object.