Based On Deep Learning Of UAV Aerial Vehicle Detection And Identification

Vehicle target detection technology refers to the process of vehicle detection and recognition from different image data sets through feature extraction.The vehicle target detection technology using deep learning method performs better in the accuracy and speed of target detection.With the development of science and technology,the detection and recognition of the vehicle from the aerial images of UAV has become an important application research direction.In this paper,we focus on the detection and recognition of UAV aerial vehicles using deep learning method,and aiming to solve the problem that YOLOv3 algorithm can't identify the incomplete vehicle targets in UAV aerial vehicles,and may not detect the vehicles in short distance.Based on these problems,we propose the algorithm for generating candidate frame of UAV aerial vehicles based on anchor points and a prediction frame of UAV aerial vehicles based on nonlinear function maximum suppression.Achievements of this paper are as follows :Firstly,propose a candidate frame generation algorithm based on anchor points to solve the problem that some incomplete vehicles can not be detected in the UAV aerial vehicles.The algorithm first optimizes the selection of anchor points,and then obtains the distance function through analysis.With this function,we can calculate the distance between the initial anchor point and the vehicle data sample point,and then obtains all anchor points.Next,we predict the moving range of anchor frame in four directions,and use sigmoid function to set its moving in a certain range.The algorithm improves the recognition accuracy of incomplete vehicles by vehicle detection network.And then,a prediction frame generation algorithm based on nonlinear function maximum suppression is proposed to solve the problem of missing detection of some vehicles which are close to each other and incline.Our algorithm selects the candidate box whose intersection ratio with the highest confidence vehicle candidate box is greater than a certain value,and recalculates its confidence through the nonlinear function.After that,we use multi-level convolution feature fusion method to optimize the confidence of candidate frame.The algorithm proves to improve the success probability of vehicle detection in vehicle detection network.Finally,train the detection model using this improved algorithm with YOLOv3 network,and detect this algorithm by using the vehicle image captured by UAV.The result of experiment shows that compared with the existing algorithm,our algorithm proposed in this paper improves the detection accuracy of UAV aerial vehicle with deep learning method significantly under the real-time performance.