Object Representation Method For Object Tracking

As an important field of computer vision research, object tracking problem is concernedsteadily in recent years. Object tracking technology for moving targets has broad applicationprospects in intelligent monitoring, military guidance, human-computer interaction, trafficcontrol, automatic navigation, behavior recognition, etc. How to select the objectrepresentation method is a critical factor for a successful object tracking algorithm. In thispaper, the object representation methods in the distribution fields (DFs) and the multi-instancelearning (MIL) tracking algorithms which are proposed recently are studied. After a detailedanalysis of their principles, implementation processes, main strengths and weaknesses, theirshortcomings are improved in the new algorithms. The main achievements are shown asfollows:(1) An object representation method with weighted distribution field is proposed bygiving a greater weight to the layers which are more discriminative to the background. Itenhances the DF’s robustness for object description and improves the performance the DFstracking algorithms. The weighted distribution field tracking algorithm adaptively weights thesimilarity of the candidate block’s and the target’s corresponding layers with correlationcoefficient. The more similar to the layers of the target and the more dissimilar to that of thebackground, the greater the weight is, and vice versa. Experiments on12test video sequencesshow that, compared to the distribution filed tracking algorithm, the multiple instance learningtracking algorithm and the compressive tracking algorithm, the weighted distribution filedtracking algorithm gets a higher average tracking accurate rate by7.17%,28.59%,6.96%incomplex scenes including obvious deformation, scale and illumination changes, occlusion androtation of the target.(2) A multi-channel Haar-like feature based object tracking algorithm with multi-instancelearning is proposed to improve the tracking performance on color videos. We propose thattarget is represented with Haar-like features generated from three channels of RGB withcompletely random location, size and channel. They represent the target using moreinformation. Next, some weakest discriminative Haar-like features are replaced with newrandomly generated Haar-like features when weak classifiers are selected. It introduces newinformation to the target model and adapts to the dynamic changes of the target appearance and external conditions. The experiment on8challenging color videos shows that theproposed method obtains a higher average tracking average accurate rate by52.85%,34.75%,5.71%than the multiple instance learning tracking algorithm, the weighted multiple instancelearning tracking algorithm and the distribution field tracking algorithm, respectively.