Object Detection Algorithm In Surveillance Videos

Object detection is an important technology in intelligent video surveillance system, is also the hotspot in pattern recognition and computer vision. The specific objects can be detected based on the analysis of videos. The task of this thesis is pedestrian detection and Pointing event recognition. Event detection has wide applications, such as intelligent video surveillance, human-computer interaction and video retrieval. The works of this thesis are as the following:Firstly, for DPM algorithm is the state-of-the-arts in pedestrian detection, we choose DPM algorithm to detect pedestrian in surveillance videos. Based on open source codes of OpenCV, we reprogram the DPM program from matlab to C++language. The speed of reprograming in C++is improved to some degree, simultaneously retaining similar detection precise with the matlab program of DPM. We have tested the detection program in the TRECVID SED videos.Secondly, we modified the detection algorithm of Pointing event of out lab, which based on human body pose estimation. We extended the features from five dimensions to ten dimensions by introduced new features.Thirdly, we implemented a detection method of Pointing event based on DPM algorithm. This method can be carried out directly to detect Pointing event in the video without pedestrian detection. If the algorithm is used based on the CNN pedestrian detection results, the detection results would be better.Finally, we presented another detection method of Pointing event based on HOG and SVM. Based on the CNN pedestrian detection, the experiments show that detection accuracy of this methods is slightly higher than the method based on DPM.