Research On The Multi-feature Fusion For Target Tracking In Video Sequences

Object tracking in video sequence is one of the most important topics in the field of computervision. The object tracking technique is widely used in fields of military and civilian entities etc,such as intelligent transportation system, video surveillance, human-computer interface andaerospace. Complex scenes, noise, illumination variation, other neighboring objects with similarappearance, object scale change, and other factors that may adversely affect the trackingperformance can not be ignored. Although a fairly large number of research achievements forobject tracking have been made in the past, there are still many problems to be solved.Using only one feature to describe the target is one of the common drawbacks for the entire targettracking algorithms. The thesis proposes a moving object tracking algorithm based on multi-featurefusion. Mean Shift algorithm’s principle is simple to understand and easy to implement, thealgorithm has a certain degree of robustness to partial occlusion of the target and backgroundchange, therefore the paper chose the Mean Shift as the basic tracking algorithm. The thesis workis focused on the key technical problems, the results and contributions about multi-feature fusiontracking technology are described as follows:Firstly, according to the derivation of Mean Shift algorithm, it is concluded that the values ofthe matching weights are the key factor affecting the tracking performance and they are related tothe Mean Shift algorithm’s iteration step. In the light of the distribution of the values of the target’smatching weights in the current frame, we defined the pseudo-moment characteristics of thematching weights. The bandwidth of the target kernel function can adjust its resize adaptively; thisstep provided more precise target information for the candidate target model in the next frame.Even if the moving target will be translation or rotation, the proposed algorithm can also adjust thenuclear bandwidth automatically. Experiments show that the new algorithm ensures the accuracyof each frame of the target model and improves the robustness the original algorithm.Secondly, the paper analyzes the movement of the target information, such as colorinformation and edge information, texture, corner features etc.. Each feature has various extractionmethods. Experiment results verified the capability of each feature in the process of the targettracking. The paper proposed several methods about the realization of various features in Mean Shift algorithm that paves the way for the third part of the multi-feature fusion. The paper alsoanalyses solution about the multi-feature fusion based on the maximum likelihood or Boost.Thirdly, according to analytical results about the capability of different characteristics in MeanShift tracking framework, the paper analyzed the performance of different feature fusion in varioustracking environments and we made a decision that more features to fuse is not simply proportionalto the beneficial effect on the tracking results. Generally speaking, two or three features to fuse areappropriate, that can reduce the complexity of the algorithm and facilitate the adjustment of thetarget tracking accuracy, and may keep the balance between the computational cost and robustnessof tracking algorithm.Fourth, the paper proposed a novel algorithm based on both the fusion of corner feature andcolor feature. First, we obtained more reliable corner information by adjusting the threshold valueof the Harris corner point detection; second, the corner point’s grayscale information wastransformed into the gradient direction histogram; third, the RGB histogram was fused with thegradient direction histogram, and the fusion histogram was used as the candidate target model.Finally, the target is tracked successfully using the Mean Shift algorithm. The experiments showthat the multi-feature fusion improves the robustness of the tracking algorithm.The technique based on multi-feature fusion not only increases the degree of the target andbackground, but also improves the accuracy and robustness of the algorithm. In the complexbackground of strong interference, the proposed algorithm remains a satisfactory real-timeperformance.