Research On Target Detection Algorithm Based On Deep Learning In Air To Ground Observation

Target detection technology in air to ground observation scenarios refers to the process of automatically identifying and locating targets in aerial images using computer vision and machine learning technology.This technology can be applied to various fields,such as national defense and security,traffic management,urban planning,natural resource management,and so on.However,aerial image target detection tasks also face unique difficulties and challenges,including issues such as the diversity and complexity of data,the accuracy and efficiency of models,the diversity and variability of targets,and the presence of a large number of small targets.In addition,to deploy aerial image target detection algorithms to drones or their vehicles,there are also constraints in terms of computational resource costs and real-time performance.Aiming at the key and difficult tasks of target detection in the aviation field,this paper selects the Vis Drone-DET dataset as experimental support.The main research work is as follows:(1)Aiming at the problem that the detection model under the convolution mechanism is not good at capturing long-distance feature information in aerial images,this paper adds a Transformer branch and FCU feature coupling module to the feature extraction network CSP-Darknet53,and proposes a hybrid backbone network with a CNN+Transformer structure.This network integrates the advantages of CNN architecture that is good at capturing local feature information and the ability of Transformer architecture to effectively grasp global feature information,And relies on the FCU feature coupling module to eliminate the semantic differences between the features extracted under the two mechanisms,thereby achieving comprehensive and accurate feature extraction.After ablation verification,this hybrid structure backbone network is superior to a single structure feature extraction network using only CNN or Transformer mechanisms in feature extraction.(2)In order to comprehensively improve the detection performance of detection algorithms in air to ground scenarios,this paper selects YOLOv5 as the reference network.Firstly,the original Backbone module is replaced by a CNN+Transformer dual branch hybrid backbone network to enhance the model’s ability to express features of targets in aerial photography scenarios;Secondly,based on dilated convolution,an adaptive coordinate attention mechanism module SKCA,is proposed and effectively applied to the feature fusion module,improving the model’s ability to locate shape-distorted targets;Finally,in view of the difficulties of small target detection in air to ground observation scenarios,this solution includes two key parts: designing a feature fusion layer and an output detection layer for small targets,effectively reducing the difficulty of small target detection at the model structure level.The results show that the detection accuracy of the proposed aerial image target detection model YOLOv5-TF in the Vis Drone-DET dataset is superior to several mainstream single stage and dual stage detection models,but the complexity of the model needs to be improved.(3)Aiming at the problem of large parameter quantity and slow detection speed of the target detection model YOLOv5-TF proposed in this paper,two improved lightweight models Slim-Yo LOv5-TF and FPGM-YOLOv5-TF are proposed based on structured pruning methods Slim and FPGM,which successfully reduce the parameter quantity of the model within the acceptable model accuracy range and improve the reasoning efficiency of the model.By optimizing and simplifying the algorithm,it can be lighter and more efficient to operate,thereby improving the feasibility and reliability of deploying aerial image target detection algorithms on unmanned aerial vehicles and their vehicles.(4)Use the AI edge computing equipment Jetson AGX Xavier to build the target detection software and hardware platform in the air ground observation scene and apply it to the airborne optoelectronic pod detection system to achieve the landing deployment of the lightweight target detection algorithm YOOv5-TF.In addition,the reliability and practicality of the photoelectric pod target detection system are verified through an application case of specific target detection tasks in urban road scenes.