Research On UAV Target Detection Algorithm Based On Convolutional Neural Network

With the rapid development of science and technology,deep learning plays a prominent role in the field of aviation,and the detection of target images by unmanned aerial vehicles(UAVs)is also very important in the process of mission execution.But now the airspace is basically saturated.Under the provisions of air traffic rules,the flight is carried out according to a certain priority.In addition,the UAV may encounter a variety of aircraft that do not comply with the regulations during the flight,or the UAV may appear "black flying" state,which easily leads to an increase in the incidence of accidents.This paper is to study on a few cases,unmanned aerial vehicle(uav)for target detection algorithm.These are the cases of a single object,multiple identical objects and multiple different objects.First of all,basic structure and the basic situation of the network are analyzed,then the above several ways separately introduced.First,the target detection algorithm of a single target is studied,and the algorithm is improved on the basis of the Faster Region Convolutional Neural Network(Faster R-CNN)algorithm,and the structure of the feature extraction part of the algorithm is improved,mainly by replacing the Alex Net Network structure of the original Faster RCNN algorithm with the improved VGG Network structure.At the same time,the selection method of bounding box is optimized to increase the optimization loss.In addition,in view of the loss function aspects proposed the false positive samples training loss,the loss function is improved.And for the training method of the whole algorithm,the gradient descent algorithm with momentum is adopted,which can be closely combined with any form of optimization.The simulation results show that the improved algorithm can complete the task more efficiently,and the accuracy is higher,up to 85%。Next to many target detection,adopting the You Only Look Once(hereinafter referred to as YOLO algorithm)algorithm based on algorithm,due to an image of the definition of a number of different targets in the sample concentration is confusion,so the study of this chapter shall not be.In this chapter,K-means clustering method is optimized based on YOLO algorithm,and Bayesian decision theory is adopted to determine the selection range of K value.After that,simulation is used to verify whether the optimization is appropriate.Simulation verification results show that the improved algorithm can better achieve the expected goal and realize target detection more quickly.Finally is to a single target and multiple targets in the mix together the target image detection algorithm,still use YOLO algorithm,and there will be no definition of hybrid multi-objective content added,due to a single target is no more training,so can separate out mixed type of the target.And will adopt mean shift clustering algorithm,clustering analysis,determine the type of k value and label,after using improved NMS algorithm for object overlap problem,on the whole structure is optimized,using the simulation,compared,the results show that the improved effect is more accurate,the speed is more quickly.