Study And Application Of Information Extraction And Annotation Technology In Surveillance Video

With the development of technology and people's awareness of safety prevention and control,video surveillance has been widely used and played an important role in traffic management,social security and other fields.Nowadays,the analysis and management of surveillance video has been gradually developed from semi manual processing to intelligent surveillance system.There are a lot of redundant information in the video data collected by the surveillance camera.How to extract useful information from a large amount of video data is an important problem to be solved urgently.It is also one of the research foucses in intelligent surveillance field,and has certain research and application value.According to the characteristics of the surveillance video,the thesis has focused on the technology of information extraction and annotation of surveillance video with the technology of object detection and tracking.The main works of the thesis are as follow:1.According to the real-time requirement of monitoring video analysis and the characteristics of monitoring video scene.In the aspect of moving target detection,the video image is preprocessed to obtain the region of interest where the moving target exists,which reduces the amount of computation and improves the real-time performance.First,we use the ViBe algorithm to extract the foreground.The method only needs one frame to initialize the model,which has the characteristics of fast speed and less resource consumption.Then the foreground image is binarized and the image is denoised by the median filter algorithm.Finally,the Canny operator is used to extract the foreground contour to get the region of interest.2.According to the accuracy requirement of video surveillance,the object detection classifier based on Haar and Adaboost is constructed on the basis of analyzing the characteristics of video image and research classifier.Mainly through the process of re-distribution of the sample image weight,training a number of weak classifier,and in the final cascade for a strong classifier for a specific target detection.In order to improve the accuracy of detection and classification of targets,the thesis has trained three kinds of classifiers for human face,human head and vehicle.In view of the problem that the constructed classifier is not efficient,the video image is copied into three layers corresponding to the target during the detection process.According to the actual size of the target size to select the appropriate scale to reduce the layer and set a reasonable range of different target size,thereby reducing the detection process generated by the amount of calculation.After experimentally,the effect of optimization is verified.3.Based on the analysis of the existing tracking methods,the compression tracking algorithm with high real-time performance is studied emphatically.The main idea is to select the image samples in different regions of the video frame,and use the sparse matrix to reduce the dimension vector of these images in multi-scale space,and then classify these eigenvectors using the naive Bayesian classifier,So as to distinguish between tracking the target image and the background image,to achieve the purpose of tracking.At the same time,aiming at the problem of inaccurate tracking caused by the change of target size during the longitudinal movement of the target in the monitoring screen,the compression tracking algorithm is improved with the SURF feature point matching.The position and size of the tracking box are corrected by using the feature matching effect between the different objects.The experimental results show that the improved algorithm has better tracking effect in the monitoring scene.4.In order to realize the semantic annotation of the target information,on the basis of analyzing the principle of semantic annotation and the characteristics and application of the special surveillance video scene,the thesis mainly studies the method of dividing color space into multiple color semantic labels by HSV model.Aiming at the problem that the information of the traditional tracing trajectory description method is redundant and not intuitive,a method of meshing the monitoring area is proposed,and the trajectory of the target is described and recorded in the form of a string.Combining the results of target detection and tracking,the color characteristics and trajectories of the target are annotated,and the validity of the method is verified by experiments.5.According to the requirements of the system and the software architecture,a prototype system for surveillance video images annotation is designed and implemented.