Research On Vehicle Target Detection Method Based On UAV Detection Video

Due to the high requirements of fixed cameras on road facilities,fixed cameras are installed to monitor vehicles on mountain roads and rural roads with no facilities to monitor the environment.Moreover,the monitoring range of fixed cameras is small and the flexibility is poor,so that all-round monitoring of road vehicles cannot be achieved.Unmanned Aerial Vehicle(UAV)have the advantages of strong maneuverability,wide detection range,and simple operation.They can assist fixed cameras to complete monitoring in areas without cameras.Therefore,the traffic video detection based on UAV shooting has high practical significance,but due to the small target size of UAV aerial photography,the current deep learning-based target detection algorithm is difficult to directly put into application.The thesis proposes a new method for vehicle target detection based on UAV video,the main researches are as follows:(1)Faced with the current shortage of UAV video vehicle target detection datasets,we determined the traffic scene near Hunan University of Technology as the construction background of the dataset,and used UAV aerial photography at road intersections,parking lots,and construction near the school.Units and other places collect the required UAV road video data sets,and through screening,a total of 14567 vehicle detection data set samples are obtained.In order to increase the labeling efficiency,a semi-automatic labeling method is proposed to label four models of car,truck,van and bus.At the same time,in order to overcome the problem of unbalanced sample data,the thesis uses random maps for data enhancement for categories with fewer data samples.(2)Due to the small size of vehicle targets in UAV videos,the detection accuracy of existing target detection algorithms is difficult to meet the requirements.For this reason,this thesis proposes an improved YOLOv4 algorithm suitable for vehicle target detection in UAV videos.Use improved k-means clustering to obtain anchor boxes suitable for UAV vehicle target datasets to improve detection accuracy;use depthwise separable convolution in the residual module of the network instead of standard convolution to reduce the amount of network parameters,and combine The convolution block attention module CBAM is used to enhance the ability of the residual module to extract features;the improved structure of SPP,ASPP,is used to increase the receptive field of the convolution kernel on the basis of ensuring the resolution.In order to further improve the positioning accuracy and bounding box regression accuracy of the UAV vehicle target,the classification loss Focal Loss and regression loss EIo U Loss are used to improve the cross-entropy classification loss function and regression loss CIOU Loss in the original algorithm.The experimental results show that the improved YOLOv4 algorithm for vehicle target detection in UAV video is better than the algorithm before the improvement,the m AP is increased by 4.4%,the F1 score is increased by 5.6%,and Check performance still has good performance in complex environments.(3)In order to enhance the human-computer interaction performance of the algorithm,the vehicle detection system was designed by calling the python graphical module wxpython,and the detection results were displayed on the interface by inlining the trained deep learning model in the GUI interface,which effectively enhanced the visualization performance of the algorithm and Convenient operability.