| Tracking moving objects is an important issue in the field of computer vision. It is widely used in military and civilian fields including video surveillance, image compression, intelligent transportation and weapons simulation training. Particle filter shows great success in non-rigid object tracking. But tracking errors would be inevitable when the background changes greatly, or when the color distribution of background is similar to target’s. And only the number of particles is enough, the algorithm can accurately track the target.To eliminate these errors, an improved tracking algorithm based on particle filter is proposed in this paper. Our algorithm integrates visual saliency information into the particle filters framework for object tracking and use Mean Shift to optimize particle collection. We extract salient region only from target area and particle area based on visual attention, and then build the target model for salient region so as to reduce the interference of background pixels. The target model is more robust than the model which is established directly based on single color because of integrating salient features in process of salient region extraction. The target model and the candidate model overcome the tracking problem of target loss which is caused by interference of similar background, illumination change and the target deformation. In the process of state evolution, this paper use Mean Shift algorithm to optimize each particle so that each particle converges to the optimum position. In this way, the algorithm can achieve better tracking performance and improve the tracking speed by a small number of particles. Further, the target template is updated adaptively during tracking so that the target template can reflect the latest feature of target. Experimental results show that the proposed method is more robust against background interference and illumination changes. Therefore, the target tracking is more stable. |
