| With the continuous improvement of the level of grain depot information,many grain depots have built a comprehensive video surveillance system covering the inside and outside of the depot,and have played an important role in grain depot management and supervision.However,the current grain depot video monitoring system has a single function,mainly recording,and its access methods are only real-time viewing and post-event callback.Video data analysis is mainly based on manual research and judgment,and the level of intelligence is low,and it requires a lot of manpower and material resources.Taking the video data center established by China Storage Grain in Beijing as an example,it manages nearly 1,000 cameras and is equipped with 6 personnel on duty 24 hours a year,with an annual operating cost of nearly 10 million.Unable to achieve full coverage of data.In addition,on the one hand,massive surveillance video data takes up a lot of storage space,resulting in a great waste of resources;on the other hand,a dedicated person is required to conduct a spot check on the monitoring content of the day every day.Unable to quickly grasp the monitoring situation of the day.Therefore,the introduction of intelligent video processing technology is an important means to solve the current standardized supervision and management of grain depots.The main work of this article is as follows:1.In terms of target detection and recognition,based on the characteristics of many small targets in the grain storage monitoring scene,based on the YOLOv3 target recognition algorithm,an improved YOLOv3-GD(YOLOv3-Grain Depot)algorithm is proposed.And re-create the data set for model training,and use the YOLOv3-GD algorithm to perform target recognition on the effective frame pictures of each camera extracted based on the KNN background subtraction method,and further extract key information such as the category and position of key targets in the effective frame pictures to achieve Efficient extraction of effective information in multi-channel surveillance video.2.In terms of multi-target trajectory matching and generation,the extracted category,location information and image feature information are used to match the targets at adjacent times,and the running trajectories of each target in the current monitoring scene are obtained and displayed.3.In the aspect of detecting abnormal behaviors of grain depots,in view of the lack of data on abnormal behaviors of grain depots and the difficulty of training the data from the perspective of image classification,a method of using human posture key point information for human motion recognition is proposed.Aiming at the problem that the original Alphapose algorithm cannot detect and recognize other key targets required in this article at the same time as the pose estimation,and the original Alphapose algorithm training process is extremely complicated and redundant,this paper first optimizes the original Alphapose accordingly and uses The optimized Alphapose acquires the key points of the human body,and then constructs the structure vector between the key points,and uses the information such as the key point coordinates,the angle between the structure vectors,and the ratio between the geometric length of the structure vector and the width and height of the human detection frame to define different The characteristics under the action state realize the recognition and classification of human actions to achieve the purpose of abnormal behavior detection and early warning of personnel.4.Developed a set of intelligent monitoring system for grain depots for visual display,simplified video playback process,and freed up massive storage space. |
PDF Full Text Request