Multi-object tracking based on improved Mean Shift

Abstract

Since Mean Shift algorithm can not track multiple objects, a full automatic multi-object tracking algorithm based on improved Mean Shift is proposed. The background subtraction image kernel density estimation algorithm is used to detect the foreground. The extracted moving objects are used as candidate template to eliminate the influence of background. By adopting object matching based on distance matrix, new objects entering to the scene and occlusion-split between objects could be handled. The tracking accuracy is increased by using shadow removal and morphology processing. The experiment results show that the proposed method can achieve multiple-object tracking accurately, and deal with the occlusion-split between objects very well.