Research On Vehicle Detection Algorithm Of UAV Aerial Image Based On Improved YOLOv4

Vehicles are an important component of the road traffic system.The collection and identification of vehicle target information is one of the effective means to alleviate traffic safety,congestion and other problems.UAVs are widely used in the research of road vehicle information due to their small size and high real-time performance.YOLOv4 performs better than the traditional image target detection effect,but the accuracy of the image target detection under aerial photography is low.The specific performance is:(1)The background noise suppression effect is not strong;(2)The gradient disappears during the algorithm training process.The problem caused the model learning effect to deteriorate;(3)The detection accuracy is low when the positive and negative samples of the data set are seriously imbalanced.Therefore,this thesis is based on YOLOv4 to improve,combined with the affine characteristics of the spatial transformation network to enhance the suppression of the image background noise,the dense connection network is used to reduce the gradient disappearance problem,and the focus loss function is used to solve the positive and negative sample imbalance problem.On this basis,research the network model suitable for UAV aerial vehicle detection,with the goal of improving the detection accuracy of the model.The main research contents are as follows:(1)Data set preparation.First,a UAV was used to collect road vehicle images near the Huanpu Science and Technology Park on the West Third Ring Road in Xi’an.After the data set was filtered,expanded and cropped,it was then labeled.Then select Vis Drone2019-DET to expand the data set,filter it and convert the data format of the annotation file.Finally,the processed image data is combined to form the data set of this thesis.(2)An improved aerial vehicle detection algorithm ST-YOLOv4 based on spatial transformation network is proposed.In view of the many noise problems in the vehicle images taken by drones,a spatial transformation network is added to the backbone feature extraction network of the YOLOv4 network model to enhance the saliency of the main features,suppress background noise,and reduce the missed detection rate and error Inspection rate.The experimental results show that the improved model ST-YOLOv4 has a vehicle recognition accuracy of 95.78% in aerial images.Compared with the original YOLOv4 detection model,the detection accuracy is increased by 2.71%,which proves that the model has a good performance in aerial vehicle target detection.Robustness.(3)An aerial image vehicle detection algorithm ST-Dense-YOLOv4 which introduces a densely connected network is proposed.In view of the disappearance of the gradient in the network training process and the generally small size of the vehicle target in the aerial data set,the problem of low detection accuracy of the model small target is caused.The densely connected network is introduced into the enhanced feature extraction network,and the feature transfer and feature extraction are enhanced through the multiplexing of the densely connected network.At the same time,the focus loss is introduced,and the sample balance coefficient and modulation coefficient are added based on the original loss function.Solve the problem of imbalance between positive and negative samples in the data set.The experimental results show that the detection accuracy of the algorithm reaches 96.75%,which is 1.54% higher than the ST-YOLOv4 network detection accuracy.It also verifies that the detection accuracy of small target vehicles is better.