Research And Implementation Of Musk Deer Behavior Detection System For Video Surveillance

Using the monitoring system to monitor the behavior of musk deer is a common technical means in aquaculture.By using computer vision technology to detect the signs of musk deer’s estrus and injury in the video,that is,climbing and fighting behavior,it can solve the existing problems of high abortion rate and high injury rate in the musk deer breeding base.Compared with the way of manually viewing the surveillance video to capture the target behavior of the musk deer,the way of using computer vision technology is more efficient.The following problems exist in the current musk deer behavior detection research:(1)the standard musk deer behavior dataset has not been established.(2)the factors such as behavior occlusion or blur make the accuracy of behavior detection low.(3)The detection speed of the video target detection algorithm is slow.Based on the above analysis of the problems existing in the behavior detection of musk deer,this paper mainly does the following research:First,a dataset of musk deer behavior was created that includes both climbing and fighting behaviors,and the dataset was enhanced through data enrichment to provide a data base for further research.Second,this paper proposes an improved model for target behavior detection based on MEGA(Memory Enhanced Global-Local Aggregation).Using the memory enhancement module in the MEGA model,the global semantic information and the local position information are aggregated into the features of the current image to solve the problem of blurred occlusion of video data.Improve the generation method of the anchor,use the algorithm GA-RPN(Guided Anchor-Region Proposal Network)to generate a highquality previous image that matches the target,reduce the amount of model computation,and improve the recognition speed.At the same time,the label matching strategy has been improved,and the SimOTA algorithm is used to dynamically calculate the number of positive samples required for the target,and high-quality positive samples are selected for matching so that the model converges quickly.To accurately evaluate the improved algorithm,ablation experiments are conducted in this paper using the selfgenerated musk deer behavior dataset and the public ImageNet VID dataset.The experiments show that the improved target behavior detection model in this paper achieves the optimal performance on the self-created dataset while reducing the detection time.Finally,based on the model proposed in this paper,a system for detecting the behaviour of musk deer is realized through an analysis of the requirements,an overall architecture and the development of functional modules.The system mainly includes four functional modules:intelligent analysis,data management,model management,and user management.In order to verify the function and performance of the system,the system is tested.The results show that the system can correctly detect the behaviour of the musk deer,display and store the detection results,manage and update the model,and realize intelligent monitoring of the behavior of the musk deer.