Visual Tracking And Description Of Pedestrian Targets In Surveillance Scene

The technology of intelligent video surveillance is gradually replacing or changing the traditional video surveillance system, which becomes the main direction of enhancing the system’s performance. Its core technology is video analysis, including detection, identification, tracking and description of the targets in the video. So it is really a vital part to track the interesting targets and generate structured semantic description of the tracked ones. At present, since that the structures of rigid bodies, such as vehicles, are relatively standardized, the corresponding methods for tracking and describing have been more mature. As a result, it is particularly important to do pedestrian-oriented research on tracking and describing for improving the existing intelligent video surveillance system.Firstly this paper analyzes the basic principles of popular online tracking algorithms and highlights one tracking method based on compressive sensing(referred to as CT). Then combined with the demand of real-time and reliability for visual tracking in actual monitoring scenarios, we study the single-target tracking algorithm by improving CT. On that basis, this paper also makes a thorough study on the main problems of multi-target tracking. Focusing on the mutual occlusion between targets, a novel method for detecting and handling occlusion is proposed to achieve locked-tracking of targets. Finally, the structured description for pedestrian targets is studied in detail, especially the pedestrian-oriented content description method based on XML description language. Detailed discussions are as follows:To solve the single-target tracking problem, this paper has introduced scale changes and linear prediction, which improves the shortcomings of the existing CT algorithm. And then based on the open-source tracking algorithms, a new platform is built to test various tracking algorithms. In addition, by designing integrating strategies of different algorithms above, we make some further improvement of the tracking accuracy.To solve the multi-target tracking problem, this paper gives a detailed introduction of the system’s main modules and their function. And then we mainly analyze the real-time problem and occlusion problem in actual multi-pedestrian scenarios. Based on the Kalman Filter and pedestrian feature classification, we have improved the previous data association method to achieve locked-tracking under severe occlusion.Focusing on the content description of pedestrian targets, this article generates the structured content description in XML format, according to the results obtained by tracking, which includes the underlying features of color, texture, shape and spatial locations. In the meantime, the description information is packaged as TS stream and transmits from client port to server port.The typical surveillance videos, such as transportation or bank scenario, and other public test sequences, such as PETS, are utilized to verify that the improved tracking and description method proposed in this article is quite effective, which provides support for processing mass surveillance video data.