| With the continuous development of urban traffic and the increase of intelligence,traffic management departments need to monitor and master the traffic situation on the road more efficiently and accurately.Vehicle location detection and traffic statistics based on UAV aerial photography and deep learning technology has become a research hotspot.Based on UAV aerial photography and YOLOX algorithm,this paper proposes an improved YOLOX method for small targets of UAV aerial photography for vehicle localization detection and traffic statistics problem.Experimental results show that the method proposed in this paper can accurately and quickly detect vehicles and count vehicle traffic,providing an efficient and accurate traffic data analysis and management tool for traffic management departments.The main contributions of this paper are:(1)Improvement of YOLOXTo address the problems of occlusion,overlap,and small targets in the process of UAV aerial photography detection,which lead to the loss of detection objects and low accuracy,this paper improves the resolution of small targets as well as enhances the extraction of small target information by adding the module of attention mechanism to the super-resolution method on the basis of YOLOX network;to enhance the efficiency when fusing between different feature layers,a new feature fusion In order to enhance the efficiency of fusion between different feature layers,a new feature fusion calculation method is proposed to improve the detection accuracy of small and medium-sized targets;a tail-end perceptual field expansion layer is designed to expand the perceptual field of the detection feature layer,so that the detection head can use more object information to locate and distinguish dense objects,thus improving the detection accuracy for occluding dense small targets.(2)Subsequent algorithm combination and improvementYOLOX is based on YOLOv3,but it is not as fast as YOLOv3 in terms of computation,number of parameters and training speed.Therefore,this paper improves the training speed by transplanting the improved YOLOX algorithm into the framework of YOLOv5 algorithm.This paper also tries various methods to improve the overall algorithm accuracy,such as using CIo U loss function instead of the original IOU loss function to improve the algorithm localization accuracy;using a more suitable category loss function calculation method for aerial photography environment,etc.;using Soft NMS algorithm to solve the model in the vehicle density problem,etc.After the above improvements,the experiments are tested on the test set of the dataset Vis Drone,and the results show that the AP50 result of the MFEYOLOX network is 47.78%,and the accuracy is improved by 9.43% with the similar number of parameters and computation as the original network.For the problem that the computational volume and complexity of the detection network are too high and unfavorable for field detection,the original network is lightened in this paper.The efficient channel attention mechanism is chosen to be added to the Ghost module,which is used as a whole to replace the CSPBettleneck in the backbone network,and deep separable convolution is used in other parts of the network,which not only reduces the complexity of the network and is more conducive to deployment in practical applications,but also enhances the information exchange and fusion among the channels in the model,making the backbone network can extract more feature information.(3)Implementation of traffic statistics and other functionsIn this paper,we use UAV to obtain road aerial photography data,and use the target detection algorithm combined with tracking algorithm after the above improved network structure to detect and count the vehicles in the video,and equipped with a license plate recognition module,which can be used to quickly locate and find the target vehicles. |
