Research On Fine-grained Object Detection Algorithm Based On Video Srteams

With the development of artificial intelligence technology,video detection has been widely used in various fields,such as road traffic,military security,public safety and so on.In recent years,in order to enhance the intelligent management level of office space,more and more units use video detection technology to manage internal personnel and visitors.The fine-grained object detection in the video stream studied in this thesis is a part of the sensorless access control system.In view of the low detection accuracy of the existing target detection algorithms on fine-grained objects and the difficulty of real-time video detection,this thesis proposes a new target detection algorithm based on the combination of optical flow method and improved YOLOv4,and applies this new target detection algorithm to the window switch detection of non inductive access control.The specific research will be carried out from the following aspects:Firstly,aiming at the problem of low detection accuracy of existing target detection algorithms on fine-grained objects,this thesis improves the boundary box loss function and prior box generation mechanism of YOLOv4 algorithm.Through experimental comparison,the detection accuracy of the improved model on fine-grained data sets is2.13% higher than that of the original model.Secondly,based on the pre-training model,the improved YOLOv4 model is trained by using the labeled data samples.After the model training is completed,the test performance of the model is tested by the given evaluation index,and the experimental results are analyzed.After many training and adjusting parameters,the optimal detection model is determined.Finally,aiming at the problem that video stream detection is difficult to meet the real-time performance,this thesis combines the optical flow method with the improved YOLOv4 algorithm.The optical flow method is used to extract the moving target area,and then the improved YOLOv4 model is used to detect and recognize the fine-grained object again,which reduces the amount of detection calculation and avoids the false detection of the static area.The experimental results show that the average detection speed of the new target detection algorithm proposed in this thesis reaches 27.95 frames per second,which is significantly higher than that of the original YOLOv4 algorithm,and realizes the real-time detection of video stream.