Research On Vehicle Detection Algorithm In Aerial Image Based On YOLOv4-Tiny

Intelligent transportation system can monitor the road condition,which is the inevitable direction of traffic development.Vehicle detection plays an important role in intelligent transportation system.It can detect the traffic flow and provide various important data for road traffic.Because the background environment of aerial images is complex,the number of targets is small and large,and the occlusion problem is serious,which brings difficulties to aerial vehicle detection.So how to improve the accuracy of vehicle detection in complex environment has practical significance.YOLOv4 algorithm represents the most advanced object detection level in the industry.Its detection speed is faster than other detection systems,which realizes the trade-off between speed and accuracy.YOLOv4-Tiny is a simplified version of YOLOv4,and its detection speed is faster.Compared with YOLOv3-Tiny,its accuracy has been greatly improved.However,there are still some areas that need to be improved in the detection accuracy of aerial vehicle.Therefore,this paper studies vehicle detection in aerial images based on YOLOv4-Tiny network.Firstly,a new data augmentation method Mixup Mosaic for image aliasing is proposed.Through this data augmentation method,negative samples are added to the training to improve the recognition effect of the model.Then,in order to solve the problem of poor detection effect of aerial vehicle small target,this paper proposes an improved YOLOv4-Tiny aerial vehicle target detection algorithm.Firstly,1×1 convolution is added to the YOLOv4-Tiny backbone network to improve the feature extraction ability of the backbone network.Then,the deep feature upper sampling is fused with the shallow feature to get a detection layer with 8 times down sampling suitable for small target detection.In order not to increase the complexity of the model,the detection layer with 32 times down sampling suitable for large target detection is removed,and the K-means++method is used for aerial vehicle detection The anchor frame is selected again to make the anchor frame more suitable for the size of aerial vehicle and further improve the detection effect of aerial vehicle.Based on the platform of single card NVIDIA geforce GTX 1060,experiments are carried out on visdorone2019 data set to verify the proposed improvement.The experimental results show that,on the basis of using Mixup Mosaic data augmentation method to add negative samples to the training,this paper completes the training of aerial vehicle detection on the improved YOLOv4-Tiny network,and tests on the test set,and the map reaches 50.17%,which is 10.14%higher than the original YOLOv4-Tiny network!It shows that the method of using Mixup-Mosaic data augmentation to add negative samples to training and the improvement of the model in this paper do improve the effect of aerial vehicle detection to a certain extent!...