Research On Object Detection Algorithm Based On UAV Aerial Image

In recent years,with the continuous maturity of UAV research and development technology,images and videos taken by the onboard camera of UAVs have been widely used in many industries.In military terms,the images collected by aerial photography using UAVs can obtain complete and comprehensive battlefield information.But how to use these images to achieve rapid and accurate military target detection is still an urgent problem to be solved under the development requirements of intelligent warfare in modern warfare.With the continuous development of deep learning,the performance of the target detection algorithm is constantly improving.However,the current mainstream detection algorithms cannot overcome the problems such as complex weather interference,target height change,and different scale in aerial images.Meanwhile,their models are large,detection speed is slow,and efficiency is too low.Therefore,this paper carries out research on battlefield target detection algorithms based on UAVs aerial images and the lightweight for the above problems.The aims are to obtain a lightweight battlefield target detection model with high accuracy,high efficiency,and low power consumption.Meanwhile,the accuracy and speed of the detection model in real combat are improved to provide fast and accurate information support for subsequent combat.The main research content and innovation of this paper are as follows:(1)Aiming at the problems of high security and lack of resources of real battlefield datasets in the military equipment industry,a method of constructing and enhancing aerial image target datasets of UAVs based on simulated real scenarios is proposed.Firstly,the original dataset of this paper is obtained by three methods: the autonomous shooting of UAV,web crawler search,and film or television resource interception.Secondly,the image in the original dataset is flipped,adding noise,fog,and other operations to realize the UAV perspective and aerial image simulation under complex weather.This method can solve the problem of single shooting direction and angle of basic image data,meanwhile,increasing the authenticity and credibility of data.Then,a random image Mosaic method is proposed to enhance the small target proportion.This method is suitable for the characteristics of wide coverage,high resolution,and small target proportion of aerial images,and can further improve the generalization of the model.Finally,a data enhancement method based on random target extraction and embedding is proposed to enlarge the proportion of tank and military-car targets.This method can solve the problem of the unbalanced target distribution and single background in the original dataset and can increase the complexity of the image background.The experimental results show that the data enhancement method used in this paper can effectively improve the authenticity and diversity of the aerial battlefield dataset,and can also improve the accuracy of model detection.(2)Aiming at the problems of low detection accuracy and high false detection rate in the background of real complex aerial image battlefield data,a high-precision aerial image target detection model based on improved YOLOv5 is proposed.Firstly,a multi-feature cross-fusion attention mechanism is proposed and introduced in the backbone network.This method can not only solve the problem of complex background of aerial images and the presence of similar objects interference but also enhance the feature interaction in both spatial and channel dimensions,to improve the detection performance of the network in complex backgrounds.Secondly,the BIFPN structure is introduced and improved to solve the problem that the aerial image targets are of different scales and the model has poor cross-scale feature fusion effect,which easily causes target miss detection.The structure first uses cross-level cascaded paths for multi-scale feature fusion to reduce the loss of shallow-level information.Then deconvolution is used to compensate for the feature loss due to nearest neighbor interpolation.And ASFF is used for adaptive spatial feature fusion to further enhance the cross-scale fusion between different targets and reduce the miss-detection rate of small targets.Eventually,the detection effect of the whole algorithm is improved.Finally,the experimental results show that the algorithm of this paper can effectively improve the detection accuracy of the model.At the same time,it can reduce the rate of missed detection and false detection in complex environments,and achieve highaccuracy detection of aerial battlefield targets.(3)Aiming at the problems of large volume and slow running speed of the improved aerial target detection network model,this paper carries out lightweight processing for it.The speed of the model is optimized under the premise of ensuring certain accuracy.Firstly,the lightweight Ghost convolutional module is used to replace the redundant general convolutional module in the original network to reconstruct the backbone network.This method can reduce a large number of operational problems caused by the original network convolutional layer and reduce the amount of model computation.Secondly,a model pruning method is used to filter and eliminate the unimportant channels in the network to solve the problem of large model size and redundant structure.And this method can further reduce the number of parameters and computation of the whole model,and improve the overall detection efficiency and speed.The experiments show that the proposed method is effective for lightweight models and can realize rapid and efficient aerial battlefield target detection.