Research And Implementation Of Multi-Scale Object Detection Algorithm In UAV Aerial Images

In the age of Artificial Intelligence,the UAV Aerial technology is becoming more and more important in logistics,rescue,security,agriculture and so on.The Object Detection technology for UAV Aerial Images is a current research hotspots in the academic circle.However,detecting objects in UAV Aerial Images poses several challenges,such as small object scale,complex background,and limited computing power,which makes it difficult to balance real-time performance and high detection accuracy.In this paper,in order to address these issues,the Multi-Scale Object Detection for UAV Aerial Image is studied,and the work focuses on three aspects:Firstly,to solve the issue of complex backgrounds and the large scale variation in UAV Aerial Images,which contributes to the much loss of feature information during the feature fusion stage,an improved CC-Yolov5 network is proposed.This involves introducing CBAM and C3 TR modules into the Yolov5 network to obtain richer feature information.Additionally,two layers are added to the Neck part to better fuse the acquired feature information.The experimental results on the Vis Drone2021 dataset show that the m AP reached to 34.2%,while the number of parameters only increased by 3.9M.The average time of detecting an image on the server was 0.023 s,which meets the requirement of real-time detection and improves the detection accuracy.Secondly,due to the low resolution of the input image,the resolution of small objects will be lower and there is too little visual information in the image,a SC-Yolov5 network is proposed.This involves adding the SOCA module to the CC-Yolov5 to improve the image resolution and enhance the feature information of small objects.Furthermore,the CBAM module has been removed to other location in order to improve the detection performance.Finally,the training and testing are carried out on the Vis Drone2021 dataset.The m AP on the Visdrone2021 dataset is increased by 13.3% on the basis of Yolov5 m,and the average time of detecting an image is 0.022 s,which greatly improves the detection accuracy and can meet the requirements of real-time detection.Finally,considering the limited computing power of Nvidia Jetson Nano,a lightweight improvement is made on SC-Yolov5.This involves replacing the original Conv module of the network with the Ghost Conv module and replacing the Shuffle Net V2 with Backbone.As a result,the parameter number of SC-Yolov5 is reduced from 24.8M to 11.1M,and GFLOPs is reduced from 58.6 to 7.5,greatly reducing the network complexity.However,the m AP is reduced by 19.2%.Nonetheless,the detection accuracy has decreased to 24.6%,but the average time of detecting an image on the edge equipment is 0.038 s,enabling it to run on equipment without delay and obstruction.