英文摘要 | Visual tracking aims at estimating the object states such as position, orientation,scale etc from the image sequences, is an important part of computer vision research. It is always a hot topic and attracts much attentions, attributing to the following two reasons. For one thing, visual tracking, an intermediate-level vision part, lays a solid foundation for the high-level activity understanding and content based video analysis. The tracking has wide applications ranging from visual surveillance, video retrieval to human computer interaction. For another, the factors such as pose and illumination variation, occlusion, fast motion, similar target distrctors etc make robust visual tracking a challenging task. With decades of developments, there emerge many tracking approaches. Yet there do exist many theoretical and technical problems, which desire our further efforts. This work makes focus on multiple target tracking (MTT), which is a subclass of object tracking. The most important difference between multiple target tracking from the single object tracking (SOT)is that MTT needs to keep the ID correct for all targets in the tracking process. ID maintaining is a tough task, especially in the case of multiple target interaction or crowd/semi-crowd scene. The essential reason for the challenge lies in the multiple association hypotheses, where one target has many associated target candidates in the neighbor frame and there exist large association ambiguities. All our work aim at alleviating the association ambiguities and making multi-target tracking robust and efficient. The main contributions of this thesis are summarized as follows: 1) Online random forest based multiple target tracking. Our approach has three advantages. First, the discriminations between different targets are highlighted by using the online random forest classifier, thus alleviating the association ambiguity naturally. Second, the classifier is very robust to temporal varied target appearance by using online learning style. Finally, the subspace representations of targets are learned and used together with online random forest. By integrating discriminative model and generative model, the multi-target tracker achieves good performance. 2) Using maximum consistence for multiple target association in wide area traffic scene. In this work, we pay attention to an important application, wide area traffic scene surveillance. First, we propose a coarse to fine vehicle detection framework,... |
修改评论