Deep Learning-based Vehicle Extraction Algorithm For UAV Remote Sensing Images

Sudden geological disasters often paralyze local transportation networks,and emergency rescue measures and construction operations of transportation networks should be met promptly when disasters occur.With the rapid development of UAV technology and remote sensing technology,new ways of real-time monitoring of transportation networks have emerged.This paper summarizes the research status of UAV image stitching technology and remote sensing image recognition technology at home and abroad,studies the feature point alignment and stitching of UAV remote sensing images,and uses convolutional neural network to classify and recognize UAV remote sensing images,in which the recognition object is the vehicle in smoke sensing images.The main research contents and research conclusions are as follows:(1)Processing UAV remote sensing images.Firstly,the SIFT algorithm based on feature matching is selected to extract feature points of multiple images and will perform the alignment operation,then the alignment accuracy between images is improved by the method of coarse difference removal,and finally the stitched image of the area to be measured is obtained.(2)UAV remote sensing image target detection and recognition.The two most widely used convolutional neural networks,the Faster R-CNN model based on region suggestion and the YOLOv5 model based on regression learning,are selected,and the two models are trained using the exact same data set to achieve the detection requirements of target vehicles.The experiments show that the average detection accuracy of the Faster R-CNN model reaches85.9%,which meets the detection requirements,but the model has the phenomenon of missing detection for targets in complex backgrounds,dim skies and their own small size.The YOLOv5 model achieves an average detection accuracy of 86.7% and is able to detect the missed targets in the Faster R-CNN model,but the detection accuracy of these targets is still low.(3)Improving the YOLOv5 model.Based on the YOLOv5 model,its network framework is improved,and the Auto-YOLOv5 model is obtained by adjusting and improving the C3 module,Swin Transformer module and feature fusion process to achieve the purpose of improving the detection accuracy and detection rate.After experimental comparison and analysis,the average correct detection rate of Auto-YOLOv5 model is about 1.6% higher than that of YOLOv5 model,and the average detection accuracy reaches 88.3%,and the detection accuracy of targets that are difficult to detect due to the interference of environmental factors is also significantly improved to meet the detection needs.In summary,this paper combines the characteristics of UAV and remote sensing images,targeted processing of UAV remote sensing images and classification and recognition of specific targets in images,and the proposed Auto-YOLOv5 model can identify vehicles in images more effectively and serve the emergency rescue of transportation networks.