Research On Surveillance Video Synopsis Technology Based On Deep Learning

With the in-depth development of the mobile Internet and the Internet of Things,surveillance cameras are not only used in public places,but are also used more and more in private places such as homes,which guarantees people’s safety.However,surveillance video data has shown an explosive growth trend in recent years,which has brought great pressure on people to store and browse video data.Therefore,the development of surveillance video synopsis technology is of great significance to reduce storage and labor costs.Although researchers have done a lot of related research on this,because video synopsis involves complex technologies in video target detection,target tracking,etc.,there are still problems such as inaccurate target detection,low video compression ratio,or serious target collision.This thesis takes surveillance video as the research object,researches and implements the video synopsis algorithm,overcomes the problems existing in the current video synopsis technology,and improves the compression ratio and viewing effect of the synopsis video.The specific work is as follows:First,through the research and analysis of traditional and deep learning target detection algorithms,the Mask RCNN algorithm is determined as the basic algorithm of the target detection module.In order to better balance the accuracy and speed of target detection,the backbone network Res Net50 for feature extraction in the original algorithm is changed to Lambda Res Nets.The experimental results show that the algorithm based on the improved Mask RCNN effectively improves the detection accuracy,and the speed is faster than the original algorithm under the condition of ensuring reasonable accuracy.Secondly,in order to solve the problem of unbalanced positive and negative samples in the target tracking field,an improved MDNet-based video target tracking algorithm based on the Multi-Domain Network(MDNet)algorithm is proposed.First,in order to enrich the positive samples,our algorithm adopts a method based on the confidence of the candidate frame to select positive and negative samples;secondly,considering that updating with the samples taken from the current frame may reduce the reliability of the model and lead to the degradation of subsequent tracking performance,the center point coordinate variance threshold is used;finally,a better Focal Loss loss function is adopted to replace the cross entropy loss function of the original algorithm.Experiments show that our algorithm can effectively improve the success rate and accuracy of tracking.Finally,the foreground images of the same moving target appearing continuously in the video frame are processed into the target tube to form a tube set,which is optimized and then spliced with the background frame to generate a synopsis video.At first,the tube data is preprocessed,and the background frame is estimated by the temporal median filtering method.Then a synopsis strategy combining time shift,reducing target size and adjusting target speed is presented,and the metropolis Hastings video synopsis algorithm is used to optimize the tube set to obtain the best tube arrangement scheme.The duration compression of the synopsis video is realized by shifting the target in time,but the simple compression time will lead to the collision of the moving target and seriously affect the viewing effect of the synopsis video.Therefore,the compression ratio and collision artifact are balanced by time shift combined with the reduction of the size of the moving target and the acceleration and deceleration of the target.The synopsis video based on our algorithm increases the utilization of time and space of moving targets in the video,improves the compression ratio,further compresses the duration of the video,improves the viewing effect of the synopsis video by using the accurate target segmentation,and greatly facilitates the storing and browsing of the synopsis video.