| In recent decades, since vision sensors have been used extensively in the field of safety precautions, it's vitally necessary to develop the intelligent visual surveillance system which can replace the traditional passive surveillance system that uses analog vision in the semi-manual mode. Most importantly, though, detecting and tracking multiple moving objects are two basic elements in the intelligent visual surveillance system and key difficult points as well. This dissertation focuses on the methods of detecting and tracking multiple moving objects. A method of detecting objects in complex scenes is presented, and three methods of tracking objects are proposed according to different tracking requirements, moreover, an intelligent network visual surveillance system is independently developped to verify these methods. Main results and contributions of this dissertation are as follows:In chapter 1, the research background of intelligent visual surveillance is introduced firstly. Then, the composition of the intelligent visual surveillance system is analyzed, and the research progresses of key technologies in intelligent visual surveillance are generalized. Furthermore, on the basis of the research, the difficult problems in visual surveillance are discussed and the research significance of this dissertation is also presented.Nearly every intelligent visual surveillance system starts with object detection, and the performance of tracking objects depends on the accuracy of object detection. Because the background is usually cluttered and not completely static in the real world, it is difficult to detect objects accurately by the traditional method of background subtraction. In chapter 2, on the basis of analyzing traditional parametric models and non-parametric models, we propose an object detection approach using Ant Colony System (ACS) in a MAP-MRF framework. The MAP-MRF framework is built by the conditional probability of scenes computed by the mixed kernel density estimation method and the prior probability of pixel labels acquired in Markov Random Field (MRF). Then, the problem of optimizing pixel labels is represented as the problem of minimizing posterior energy function in the MAP-MRF framework which is implemented by the ACS algorithm. Consequently, multiple objects are detected efficiently according to the optimal pixel labels in the current frame. Extensive experiments show the utility and performance of the proposed approach. After motion detection, surveillance systems generally track moving objects based on features extracted in the detected regions of objects. Because the appearance of objects is usually different, the appearance model is a traditional method for tracking multiple persons, but the main difficulty is to represent appearance reliably and effectively, especially in presence of occlusions. The traditional appearance model containing color and motion features is introduced in chapter 3, and then, an effective tracking algorithm based on attributed relational graph (ARG) is used to track multiple persons even under occlusions. The appearance of each person is expressed by an ARG model which not only combines color feature with spatial information but also illustrates the relations among body parts. The similarity of ARG models is computed to build a matching matrix in consecutive frames. Four tracking situations are determined according to the matching matrix. In addition, in order to track persons under occlusions, probabilistic relaxation labeling in the ARG models of body parts is deduced to label occluded persons optimally. Experimental validation of the proposed tracking method is verified on indoor and outdoor sequences.In chapter 4, a novel algorithm is proposed for tracking multiple objects even if the regions of detecting objects are incomplete. According to color and spatial features, the tracked objects in the previous frame and the foreground regions in the current frame are divided into 'object patches' and 'foreground patches' respectively. We consider the problem of object tracking as the problem of optimizing labels of foreground patches. Attributed relational graph (ARG) is employed to describe appearance and structural features of object models and foreground patches, then, the optimized objective function is formed by the matching degree of these ARGs which is solved by probability relaxation. The genetic algorithm is used to compute the objective function to get optimized labels of foreground pixels, therefore, multiple objects is recognized and tracked successfully according to these labels. The experimental results perform suitably in several challenging image sequences with less foreground accuracy, which show that the proposed approach is promising.It is difficult to solve the problem of data association if objects with little distinguishable features are tracked in large-scale monitoring scenes. In chapter 5, we present a method of tracking multiple objects in real-time based on improved Joint Probabilistic Data Association (JPDA) which only uses objects'motion features. The k-best joint events are computed by the simplified murty algorithm which reduces the complexity. Data association is handled according to the association probability of JPDA even when objects enter and exit the field of view, merge and split (objects are detected as fragments). The experimental results are obtained on the standard video databases. It is shown that the method realizes tracking multiple low-resolution objects effectively in real-time and the performance is improved more greatly than the traditional JPDA method.In chapter 6, an intelligent network visual surveillance system is developed, and the software and hardware structure of this system is introduced. This system provides us ideas to present these methods in this dissertation, and it can be used as an experiment platform to verify methods.Finally, the main innovations of the dissertation are summarized, and the fields for further research are prospected at the end of the dissertation.
|
PDF Full Text Request