Research On Object Detection Of Surveillance Video Based On Deep Learning

In recent years,the wide application of video monitoring system makes it play an increasingly important role in disaster warning,security,transportation and other fields.However,in video monitoring,the amount of video data is large,and the efficiency of extracting image information by manual is low.With the development and improvement of deep learning theory,it can improve the efficiency of information detection in video surveillance and get more valuable information.However,the deep learning algorithm requires a lot of computation,parameters and hardware performance.In view of the difficulty of deep learning algorithm to meet the real-time and accuracy requirements of video monitoring,this paper studies the clustering and network compression methods in the model of YOLO algorithm,and designs the classification algorithm of monitoring video moving target.The main work of the thesis is as follows:(1)In this thesis,a bisecting K-means++clustering algorithm is proposed for target frame clustering under the framework of YOLO.Based on the detailed analysis of the implementation principle of the YOLO framework and its K-means clustering algorithm,the bisecting K-means++clustering algorithm is designed,and the longest distance criterion between the initial clustering centers is designed based on the probability of the roulette method.Using VOC2012 data set to analyze the binary K-means++clustering algorithm,the simulation results show that the binary K-means++clustering algorithm is 2.5%higher than the K-means clustering algorithm of the YOLO model.(2)The compression algorithm of YOLO model based on pruning method is designed.Pruning algorithm prunes the network structure and uses L1 regularization sparse BN layer scale factor to realize network compression.The simulation results of VOC2012 image data set show that when the pruning ratio is in the range of 30%～50%,the forward computing time of YOLO algorithm model after network compression is reduced to 57.68%～83.33%,the parameter amount is reduced to 70.08%～83.41%,and the map is reduced by 3.02%～4.33%.(3)This thesis proposes a new classification algorithm of surveillance video moving target,which combines the improved YOLO algorithm with the mixture Gaussian background model algorithm.The moving target detection algorithm based on the mixed Gaussian background model is applied to realize the location of the moving target area,and the improved YOLO algorithm is responsible for identifying the target,Based on this criterion,the moving target classification and location of surveillance video are realized.The designed video moving target classification algorithm is applied to the moving target detection in a reservoir monitoring video.The result shows that the moving target with more than 70×80 pixels can be classified and located correctly,and the map reaches 0.789.