Research On Object-based Surveillance Video Summary Generation Algorithm

In recent years,monitoring video has gone deep into all aspects of human life,played an increasingly important role in the social security and other aspects.At the same time,how to store the massive and unstructured video data efficiently and scan it easily has become new research direction of researchers.In this thesis,the monitoring video digest generation algorithm is a kind of method,which can not only save effective information in monitoring video,but also store and scan them.First,we described the significance and background of this subject,as well as the research status at home and abroad.The related technologies were introduced in detail,including the image preprocessing and the moving target detection and tracking.Then,the motion target detection algorithms based on background modeling,the moving target tracking algorithm based on multi-feature fusion and Kalman prediction,the optimization of the moving target trajectory based on the energy function and genetic optimization were analyzed deeply in turn.Aiming at the common defects in motion target detection algorithm,an improved ViBe algorithm based on spatiotemporal gradient was proposed.The detection radius was adjusted through the mean of the spatial gradient,which would improve the detection accuracy.The shadow was detected with the method of edge detection and image binary segmentation,which could remove light information and reduce the interference of the shadow.Multi-frame difference was used to adjust the threshold,which could suppress wrong test caused by the background changes.According to the changes of three consecutive images,the background changes could be detected and the rapid elimination of "ghosting" was achieved.Based on the detection results of moving objects,a multi-target tracking algorithm based on multi-feature union and Kalman prediction was proposed.Firstly,the similarities of the position feature,the color feature and the SIFT feature of the moving target were combined,which could achieve the multi-objective tracking and matching.At the same time,the research model of "object" was proposed,and the detailed definition of "object" was given.Aiming at the target occlusion problem in the video,Kalman filter was used to predict the target position in the occlusion process,and finally the multi-target was completely tracked.Aiming at the problem of spatial redundancy in monitoring video,a moving target trajectory optimization algorithm based on energy function was proposed.The energy loss,the conflict energy cost and the timing consistency cost in the energy function were explained in detail.The relative optimal solution of the energy function was obtained by genetic algorithm.Finally,the background image is merged with the target image based on the Gaussian distribution method to generate a video summary.The experimental results showed that the algorithm had achieved good results.