Along with the rapid development of socio-economic, crowd and congestion in large-scale public places are rising. It is timely and meaningful to implement intelligent crowd surveillance and crowd management. The work aims at analyzing the density distribution and volume statistics of crowd in surveillance videos, which are crucial for intelligent crowd management. In view of the limited performance of normal surveillance apparatus in actual applications, the analysis algorithms should have low time complexity, to enable real-time processing. A database of crowd surveillance named as "CASIA-Crowd" is described firstly in this paper. It is divided into two parts: the crowd density database and the crowd volume database, both of which are constructed from actual surveillance videos. An uniform measurement of crowd density is proposed, together with the corresponding labeling method. Multilayered labeling is executed on pedestrian individual and volume data. As for crowd density estimation, a novel method based on the texture analysis of "image patch" is proposed. A set of image patches are generated from the video frame, and then local density level is determined by texture analysis and statistical classification. Finally, information of local density is synthesized, formulating the final result of total density. The method does well in generalization and noise-anticipation. Two novel texture features are proposed: ALBP and GOCM, as well as a texton composition description based on the bag of words model. All of the above are designed to depict the intrinsic attributes of crowd texture. Considering that crowd density is "gradient" between neighboring levels, a multi-category classification method that based on confidence analysis and statistical learning is proposed. The method combines several binary classifiers together, by binary tree based ECOC encoding and channel transmission based decoding. The method not only increases accuracy but also improves the generalization performance. Crowd volume counting is accomplished by pedestrian detection and tracking. During the detection stage, firstly, the information of motion and corner is utilized to make a preliminary estimate of pedestrian's location; and then the sliding window search based on the appearance template of human is performed, to locate objects precisely. The two-stage searching strategy not only improves efficiency of detection, but also reduces false alarms. During the tracking stage, fe...
修改评论