Research On Multi-Objective Visual Supervision Technology Based On Object Detection And Action Recognition

In recent years,our country's economy has continuly grown and promoted the development of infrastructure,production and manufacturing.But at the same time,the situation of safety production in our country has become increasingly severe,and safety accidents have continuly occur.Among them,unsafe behavior of people is one of the main reasons of safety accidents,including workers' not wearing uniforms and helmets,workers' dangerous behaviors during construction,such as making phone calls,falling down,and squatting for a long time.At present,how to reduce the occurrence of safety accidents and strengthen the safety management of workers has become a complex and important task.This paper combines the deep learning and machine learning technology in artificial intelligence to intelligently analyze and recognize a large number of collected videos,obtaining abnormal information in the image,and making timely warnings to achieve true visual supervision and overcome traditional surveillance.Some disadvantages of huge human resource consumption and untimely monitoring and management.The main work of this paper is as follows:(1)A video image collection terminal was built based on a large zoom web camera and a motorized pan-tilt to collect RGB videos of the target areas in real time.(2)Aiming at the problem of single function and low efficiency of traditional monitoring,this paper built a multi-objective visual monitoring system based on object detection and action recognition,and realized the software design of the visual monitoring system in client-server model.The construction of the tf-pose network and the program design of the multi-objective visual tracking module in the client were completed.Completed the program design of the object detection module and the abnormal action recognition module in the server,and use the network communication technology to realize the data exchange between the two ends.(3)Considering that when there are many people in the surveillance images,it is difficult to distinguish the problematic people,this paper built a multi-objective visual tracking algorithm based on Deep Sort.First,a method of generating a character box based on the human bone information extracted by the tf-pose network is proposed;the Deep Sort algorithm will track the state of the character box in the images in real time,and different IDs for each character were set,which can quickly lock the existence of wearables based on the people with irregular and abnormal actions.Finally,different people's images are extracted according to the size of the person box,and different people's images and corresponding bone information and ID information are input to the server for identification.The server will feedback the problematic character information back to the client.(4)Aiming at the problem of non-standard wearing of uniforms(protective clothing)and helmets,this paper built a Yolov4-Mobile Netv3 object detection network based on the Keras deep learning framework.The backbone network of Yolov4 was replaced with a lightweight Mobile Netv3 network,comparing to the network which has been optimized,the detection speed is improved when the recognition accuracy meets the requirements.The requirement for computer performance is reduced.Use the network to detect the images of the people after image processing,and analyze the prior knowledge of the detection results.Finally get the situation that the target person is wearing a uniform and a helmet.At the same time,in order to train the model,based on the existing images processing technology,this paper proposed a method to construct data sets of helmets and uniforms.(5)Aiming at the problem of some dangerous actions of personnel during construction,this paper built an abnormal movement location model and an abnormal movement recognition model.The location of abnormal actions is implemented based on the SVM model,which is responsible for locating the video frames where the people may have abnormal actions,eliminating the normal actions in the video,and reducing the consumption of system resources.As the abnormality is identified,the 10 frames skeleton information are input into the Bi GRU abnormal motion recognition model based on the attention mechanism after data processing to identify the type of abnormal motions.At the same time,in order to train the model,this paper constructed a data set required by abnormal action location and recognition,and proposed a method for preprocessing bone data,which effectively improved the efficiency and accuracy of model training.