Deep Learning-Based Vehicle Detection For Remote Sensing Images

Through intelligent management of road traffic,timely grasp of real-time and comprehensive road vehicle quantity,can effectively solve the current traffic congestion phenomenon,and then build a modern intelligent transportation system,of which the key technology is the accurate identification of vehicle targets.In recent years,high-resolution remote sensing images based on satellites and UAVs are often used in various target detection tasks because of their rich feature information.Compared with satellite remote sensing technology,UAV remote sensing technology is less affected by weather and has the advantages of simple operation and small size,so this thesis starts the research on vehicle detection tasks based on remote sensing images taken by UAV.Unlike the natural scene images,the remote sensing images captured by UAVs have the problems of small size and dense distribution of vehicle targets and complex background of vehicle targets,and the existing detection algorithms do not bring good detection results.To address this problem,this thesis conducts research on two aspects of detection accuracy and lightweight of detection algorithms,aiming to improve the performance and efficiency of detection algorithms,and the main research contents are as follows:（1）Aiming at the characteristics of complex backgrounds,large number of small vehicle targets and dense distribution in remote sensing images captured by UAVs,the classical algorithm YOLOv5s in the field of target detection is improved,and the dense small target detection algorithm LSA＿YOLO in complex backgrounds is proposed.In the feature extraction layer of YOLOv5s,a lightweight multi-scale feature extraction module is constructed to enhance the feature extraction capability of the network;design hybrid domain attention module that relies on multi-scale contextual information to make the algorithm focus on target information by suppressing the interference of complex background.In the feature fusion layer of YOLOv5s,the adaptive feature fusion structure is designed to reasonably allocate the fusion weights of shallow features and deep features;the hybrid domain attention module and adaptive dynamic fusion structure are applied to the PANet structure to improve the dense small target detection capability of the model in complex backgrounds.Finally,the Focal-EIOU loss function with faster convergence and better localization effect is introduced instead of CIOU loss function to further improve the model detection performance.The experimental results on the Vis Drone2021 public data set show that when the input image resolution is set to 1504×1504,the m AP₅₀ of LSA＿YOLO can reach 57.6%,which is 6.1%higher than that of YOLOv5s,which proves that the LSA＿YOLO algorithm can improve the accuracy of the vehicle target.Detection accuracy.（2）To address the problem that the LSA＿YOLO algorithm has high detection accuracy but a more complex model,a lightweight study is conducted based on the knowledge distillation algorithm to obtain the student network YOLOv5s-KD with a lighter model while ensuring the model detection accuracy.First,the baseline algorithm YOLOv5s of LSA＿YOLO is improved by introducing depth-separable convolution and removing some structures to obtain a sufficiently lightweight student network that has a basic structure similar to YOLOv5s.Secondly,LSA＿YOLO is used as the teacher network to design the knowledge distillation algorithm guided by the attention mechanism for the problem of low attention to foreground pixels during knowledge transfer,and to guide the student network to learn the attention graph distribution and feature information distribution of the teacher network.Finally,although higher attention is given to foreground pixels in the knowledge distillation process guided by the attention mechanism,global information is ignored.To address this problem,a knowledge distillation algorithm based on global feature extraction is further designed to guide students’network learning the global relational information of the teacher network.The experimental results on the Vis Drone2021 public data set show that when the input image resolution is set to 1504×1504,compared with the teacher model LSA＿YOLO,the parameter amount and calculation amount of the student model YOLOv5s-KD after knowledge distillation are reduced by 4.773M and 87.8G,the number of frames transmitted per second has increased by 95,and m AP₅₀ has only decreased by 2.2%,which not only ensures the accuracy of model detection,but also reduces the complexity of the model and improves the detection speed of the model.The thesis has 47 figures,15 tables,and 83 references.