Surveillance Video Summarization Based On Motion Human Detection

Along with the social economy and the development of computer technology, people’s public security awareness has been strengthened. For the purpose of recording and tracking sudden emergencies which imperil the safety of public better, a large number of surveillance cameras are installed in various public areas. The surveillance video can be a decision basis for the follow-up tracking. However, tens of thousands of monitoring equipment work all the time, resulting in a lot of surveillance video data. Only using the traditional approach like fast forward and rewind to find the key events in the massive surveillance video data will be a time-consuming and labor-intensive work, which looks like looking for a needle in a haystack. More importantly, it maybe miss the important information. Therefore, the urgent problem that must be solved is how to extract the surveillance video summary efficiently. The surveillance video summary technology not only shortens the original surveillance video length but also extracts the main video content accurately.For the case of close monitoring by monocular static camera, this thesis studies the surveillance video summary technology based on motion human detection. Due to the analysis of motion object extraction, motion object judgment, video shot boundary detection and key frame extraction, we gain a static video summary method which includes abundant video content and is capable of expressing semantic meaning of the original surveillance video to some extent. This thesis mainly research content include motion human detection and video shot boundary detection. The main working as follows:(1) We improve the traditional human detection method using HOG feature. Based on the basic characteristics of HOG feature and LBP feature by PCA dimensionality reduction in motion area, we propose an efficient motion human detection method. Different from the traditional human detection based on HOG feature, in this thesis, the human body detection method is aimed at the moving human in the video. The proposed method uses frame difference method combined with background subtraction method to extract the motion object firstly and then judge the motion object, which can reduce the scanning scope of human body detection. This method solves the problem of partial occlusion and time-consuming in traditional human detection based on HOG feature, improving the accuracy and real time of human body detection.(2) The detection method for single moving human is proposed by combining aspect ratio of the bounding rectangle and skin color distribution of moving targets. The method combining the frame difference and background subtraction is employed to extract moving objects. According to the aspect ratio of bounding rectangle, we divide objects into two classes:the suspected objects of single moving human and others objects. We judge whether the suspected objects of single moving human are human body or not by the skin color distribution. Those suspected objects out of the skin color distribution will be detected by method in (1). On the premise of guarantee the accuracy, the proposed method reduces the complexity of human body detection.(3) The SVM and block histogram in HSV space model is introduced into the technology of surveillance video summarization based on mutual information. Firstly, we divide video frame that is calculated in HSV space model into unequal M*N sub-blocks assigned different weights. The characteristic value of histogram of each sub-block is obtained, and the histogram of video frame is calculated by the weighted average of all blocks according to the different weights of each sub-block. Then the feature set constructed by the mutual information between adjacent frames will be used to train a SVM classifier. This method can complete video shot boundary detection and overcomes the drawback that traditional method employing mutual information to segment shot needs to set threshold manually. At last, for the video frames in the shot, we get the key frame which can represent this shot in terms of the information entropy and remove redundant frames combined with mutual information. The methods in (1) and (2) can extract the video frames containing human bodies, and we extract the surveillance video summarization in sequence of video frames including human objects by combining block histograms, mutual information and SVM. The surveillance video summarization extracted by this method can well express the original video content with less redundancy and without manual intervention.