| Many public service sites as the external window of a company or enterprise,the service quality of its staff and the standardization of the environment directly affect the external image and customer evaluation of the company or enterprise.This thesis is based on the video surveillance data in the service scenario of the electric power business hall,and through applying the massive video image data in the semi-supervised learning object detection algorithm based on deep learning to achieve the metacapabilities of semi-supervised learning,object detection,image classification,behavior recognition and other service supervision scenarios,and building an informative and intelligent service supervision system.The system can discover relevant non-standardized records and behaviors in the video data,reduce the cost of manual supervision,improve the efficiency of supervision,form a long-term supervision mechanism,and improve the service level of public service affairs,bringing significant economic and social benefit improvement.Irregular record identification and behavior recognition are the core functions of an intelligent video service supervision system,and object detection and behavior recognition are two important basic algorithms,and the main algorithm contribution of this thesis also lies in the improvement of object detection and behavior recognition methods.This thesis first proposes a novel semi-supervised learning method CFM consistent filtering mechanism,which can effectively improve the effect of the object detection model,alleviate the over-reliance problem of the current deep learning for the image labeling task with time-consuming and labor-consuming characteristics,and can make full use of the existing massive electric power business hall video surveillance image data for the optimization of object detection algorithm iterations,fully The effective value of the unlabeled data is fully explored.Through detailed comparison experiments prove that CFM in the business hall scene data set and the public data set on the single-stage object detection method and two-stage object detection method have a little degree of effect improvement.On this basis,this thesis further proposes AMA attention mask data enhancement method to alleviate the problem of low recognition accuracy of the target object work sign accounting for a smaller part of the overall video image area in the business hall service supervision scenario,so that the object detection model is more concerned with the pixel location area of interest,and tends to ignore the redundant information in the image,which can effectively improve the effect of detection model on small objects.YOLOv5 is selected as the benchmark model,based on the characteristics and distribution of the video surveillance images in the scene of the electric power business hall,to obtain a higher accuracy object detection algorithm model.This thesis also proposes a comprehensive scheme to complete the behavior recognition in the business hall service supervision scenario through 2D-image based key object detection and video understanding based behavior recognition method.And proposed MDA multi-pooling feature descriptors aggregation method to obtain a global feature description with better information expression,which solves the current video sequence behavior recognition classification network in a single pooling led to missing information issues of the global feature description,and through MDA2D and MDA3D different pooling structures are experimentally validated on two major behavior recognition datasets,HMDB51 and UCF101,and the experimental results shows that the proposed MDA method can alleviate the network overfitting problem and effectively improve the accuracy of the TSM for video sequence behavior recognition classification network with the addition of a small amount of computation.In the above object detection and behavior recognition algorithm innovation work based on this thesis designed and developed a complete intelligent video service supervision system as a carrier of all algorithm capabilities,to implement object recognition and supervision functions in specific scenarios,including the electric business hall of the opening and closing of the door recognition,staff and work sign instructions compliance discrimination,working time record recognition,service personnel behavior specification check,and applied to the actual business of the electric power business hall scenario service supervision. |
