In the past few decades, a lot of algorithms for object detection, tracking and people counting have been proposed in the computer vision field. And a lot of mature technology has been used in railway stations, airports, docks and so on. However, video analysis is still far away from satisfying people's demand for intelligent life. That's because on one hand, the development of basic algorithms is slow. On the other hand, requirements of application for real-time and robustness limit the use of many complex algorithms. The result is that applications of video analysis are still limited to the security purpose. Thus, this thesis focuses on key technologies in video analysis for application and explores how to use them in intelligent management. The main work and contribution are as follows: 1) This thesis comprehensively analyzes key technologies of video analysis in applications, as well as their advantages and disadvantages. This thesis also sums up the general framework and future direction of intelligent video surveillance systems. 2) In the scene of meeting, people varies in appearance and gestures, which makes traditional methods for people counting difficult. Considering one-to-one correspondence of participant and seat, this thesis proposes a people counting method based on a coarse-to-fine empty seat detection strategy. Firstly, the coarse classification module is used to retrieve completely empty seats. Then in the fine classification module, the contour feature and the texture feature are combined together to solve the problem of occlusion. In this process a simplified HOG feature is proposed to speed up feature extraction. Experimental results demonstrate that the proposed people counting method achieves good results in realtime. 3) For methods using the area of foreground as feature, low precision mainly results from perspective distortion. This paper presents a perspective calibration method without using reference. The perspective distortion is decomposed into two directions: a horizontal one and a vertical one. Horizontal perspective distortion is caused by different distance of points on the ground to the camera, and calibration can be achieved by weighting pixels on the ground. The weights can be studied by Gaussian Process Regression model. Vertical perspective distortion is caused by the vertical height of people, and calibration can be achieved by weighting pixels in foreground blobs. Experimental results demonstrate that the probl...
修改评论