Appearance Modeling For Visual Object Tracking

Visual object tracking is an important topic in the field of computer vision and pattern recognition. It finds numerous industrial applications, including intelligent transportation, video surveillance, aided navigation, human computer interaction, etc. Generally, a typical visual object tracking system is composed of three components: appearance modeling, object state searching and appearance model updating. Among these components, appearance modeling is the most essential problem which directly influences the tracking performance. Although a large number of appearance modeling methods for the visual tracking have been proposed for decades, it is still a challenging problem to design a robust appearance model to meet the needs of the complex tracking environment in practical application. This thesis focuses on the study of appearance modeling for object tracking in the aspects of multi-features fusion accuracy, occlusion detection and handling, structured discriminant model building and effective appearance model to distinguish different objects in multiple objects tracking. The specific research contents include:Appearance modeling based on multi-features fusion. Multi-features fusion is a useful method for appearance modeling in visual tracking. The key problem of multi-features fusion is to accurately evaluate the description ability of object of different features, base on which the tracker could adaptively adjust the feature fusion weights in the tracking process and improve the robustness of the appearance model utilizing the complementarity between features. To improve the fusion precision of the multi-features fusion method, an appearance modeling method using adaptive multi-features fusion based on local kernel learning is proposed. The tracker divides object into multi-level image block and extracts background weighted feature histogram on each blocks, and then calculates the features fusion weight for each block by using a of multi-kernel learning method. Our method emphasizes the ability of features to describe the local area of object, which can improve the fusion precision of the overall appearance model. Experiments show better accuracy and stability of our method in comparison with the traditional multi-features fusion method.Occlusion detection and handling in visual tracking. Among all the challenging factors in visual tracking, occlusion is one of the most critical issues since it is difficult to be modeled while can greatly influence the tracking result. In order to deal with occlusion, we introduce a local sparse representation model to model the occlusion, and use a kernel weighted sparse coefficient vector to detect occlusion. Based on results of occlusion detection, we exclude the occluded area in the process of computing the observation likelihood of candidates, which improves the tracking accuracy, and recover the occluded pixels in the process of observation model updating, which reduces the drifting problem.The experimental results show that the proposed method shows good robustness in dealing with tracking sequence with occlusion.Appearance modeling based on sparse coding. The previous method introduces a local sparse representation model to model the local image of the object, which gets good occlusion detection performance. However, this method only focuses on how to effectively describe the local image characteristics, while ignores the spatial structure information among these local characteristics and their discriminative ability, which offers important information for object appearance modeling. To deal with these problems,we propose a sparse coding based discriminative appearance model. Three steps are included: Firstly, a discriminant basis function constructor is put forward, with which the sparse codes of the local image patch could effectively distinguish between object and background. Secondly, concatenate all these local sparse codes together to represent the object, and present a feature dimension reduction method based on dimensional discriminate analysis. It filters the noise in the sparse codes, thus further enhances the interclass separability of the overall feature vector. Lastly, formulate tracking as a binary classification problem. The candidate with the largest classification corresponding is taken as the tracking result. Experimental results on challenging sequences demonstrate the effectiveness and robustness of the proposed algorithm.Appearance modeling method for multiple objects tracking. On the basis of researches on single-object tracking, we extend our study to multi-objects tracking task.Compared with the single one, multi-objects tracking is more complicated. Besides the conventional challenges in single-object tracking, the frequent interaction and occlusion between multiple objects also increase the difficulty of multi-objects tracking task. Unlike most previous approaches which only focus on producing appearance models for all objects, we further consider discriminative features for distinguishing difficult pairs of objects. Firstly, an SRC based global discriminative appearance model is designed for discriminating all objects. It formulates tracklets association as an SRC problem. A discriminative dictionary learning approach is introduced, which improves the SRC classification performance. By this way, the global discriminative appearance model can distinguish different objects more effectively. Secondly, an MFH based pairwise appearance model is designed for differentiating specific close-by tracklets pairs. Thirdly, a heuristic algorithm is introduced to search the optimal solution. Considerable performance improvements are shown on challenging data sets, particularly in metrics of trajectory fragments and identity switches.