| With the development of the country,hoisting operation has occupied an indispensable and important position in the construction process of large-scale projects.However,the number of accidents and casualties caused by the accidental intrusion of non operators and the lack of command personnel on the construction site also rise.Therefore,it is particularly important for personnel safety tracking and detection in the lifting operation scene.At present,the video monitoring in the lifting operation scene mainly adopts manual method,which has the problems of low efficiency and failure to alarm in time.In recent years,due to the powerful feature extraction ability of deep learning in processing images and videos,it provides an efficient and sustainable detection method to solve the above problems,and is widely used in the field of intelligent security.However,considering the interference of monitoring scene complexity,video stream quality and occlusion on tracking and detection,the algorithm needs to have better and stable performance.Based on the analysis of the existing multi-target tracking and detection theory,this paper designs a multi-target personnel tracking method for the lifting operation scene.The main work is as follows:1.Aiming at the problem of accidental intrusion of non operators in hoisting scene,a target tracking and detection algorithm based on deep learning and filtering method is proposed.Firstly,the monitoring video of each hoisting operation in southwest oil and gas field is collected to make the data set,and the detection algorithm based on Yolo v5 and deepsort is proposed.Then,for multi-target tracking in the case of low video definition,an image filtering model based on dark channel a priori algorithm is constructed to improve the definition.Finally,aiming at the shortage of video memory resources of mobile monitoring equipment,a multi video stream image parallel detection algorithm is designed to realize multi-channel video real-time detection.The experimental results show that compared with the similar tracking and detection methods with low definition,the average accuracy is improved by 0.77%.By comparing the recognition accuracy of multi-threaded recognition and image parallel detection methods,it is proved that this method can effectively save video memory resources on the premise of little loss of detection performance.2.For the multi-target tracking and detection of commanders and ordinary operators among operators,this paper designs the human trunk recognition algorithm based on open pose,classifies the identified clothes using the Resnet,and then uses the deepsort algorithm for tracking.Conduct data enhancement processing for the imbalance of the sample number of commanders in the self built data set.For the problem of poor classification effect caused by the interference of redundant information in the identified trunk area,the spatial attention mechanism is introduced to improve the Resnet.By comparing the tracking method using the original Resnet with the improved tracking method,the average accuracy of this method is improved by 1.07%.At the same time,compared with the original Yolo v5 and other mainstream models,it proves the real-time and effectiveness of the improved method.3.Based on the above improved algorithm,an intelligent platform system for safety detection of hoisting operation is designed and built.The system adopts the framework of spring boot + spring MVC + mybatis to verify the stability and practicability of the platform in the laboratory environment and the actual production scenario.The experimental results show that the platform can effectively detect abnormal conditions in the real production scene,and send real-time early warning to relevant safety responsible persons to improve safety efficiency.Record the abnormal conditions and generate reports regularly to facilitate the relevant person in charge to specify the safety construction plan,and effectively realize the intelligent safety detection. |
PDF Full Text Request