Research On Image Detection Method Of Aerial Vehicle Based On YOLO V3

With the development of science and technology and the improvement of living standards,target detection technology has provided great convenience for people’s clothing,food,housing and transportation,especially the development of target detection algorithms based on deep learning.Important application.Aerial vehicle detection technology is a hot spot in the field of target detection.It plays an important role in traffic management and accident handling,as well as in combating illegal and criminal crimes.However,due to the complexity of the detection scene and the small size of the detection target,it has brought a great test to the target detection technology.This article takes aerial vehicle detection as the object of research,uses the YOLO V3 algorithm to study it,and compares it with several commonly used target detection algorithms.Combining the detection accuracy and real-time measurement,the YOLO V3 algorithm is improved.The specific work is as follows:Aiming at the problems of low detection rate and high false detection rate in small target detection,an improved YOLO V3 aerial vehicle target detection algorithm is proposed.The idea of feature fusion in the YOLO V3 algorithm and the three-scale network for target detection are all outstanding for the detection of large and small targets.First,the network of YOLO V3 was improved.Through the downsampling operation with a step size of 4,a deeper layer of features was fused to establish a finer-grained small target detection layer.Then,a mathematical method for the reasonable allocation of anchor boxes was proposed,making The anchor box generated by K-means clustering can be reasonably allocated on three detection scales.This method has achieved good results on the VEDAI public dataset.Aiming at the problem of small target detection,a specific vehicle retrieval algorithm based on YOLO V3 aerial image is proposed,which can retrieve the corresponding target vehicle according to the set conditions.First,it optimizes its feature extraction network and optimizes it into a two-way residual module on the basis of the original residual module to improve its ability to detect small targets.And considering that the position deviation of the prediction frame in the actual detection will have a greater impact on the detection result,the GIoU position calculation index is adopted to improve the sensitivity of the network to position information.At the same time,retrieval conditions such as vehicle type and color are added to the detection,so that the model can complete the retrieval of specific vehicle targets.Experiments show that this method has certain feasibility for the detection of aerial vehiclesand the retrieval of specific vehicles.Aiming at the problem that the traditional horizontal rectangular candidate frame cannot predict the direction of the target,a directional detection method for the position of the aerial vehicle is proposed on the basis of the YOLO-E network,and the oblique rectangular candidate frame is introduced to increase the angle information,Discarding the general horizontal rectangular candidate box,reducing the interference caused by the complex background,and then constructing a multi-task loss function with target direction loss.Experiments show that this method can effectively predict the direction of the vehicle.