Research On Target Detection Algorithm Based On Drone Shooting

As the application fields of unmanned aerial vehicle(UAV)become more and more extensive,the application scenarios of target detection under UAV shooting will also become more popular.As one of the important branches in the field of target detection,exploring the target detection algorithms under UAV shooting has been a hot topic in the research area.This paper mainly studies and improves the target detection algorithm under UAV shooting,which are divided into the following five aspects:Firstly,for the target detection method under UAV shooting,we compare the traditional target detection method with the general deep learning method,and make some comparative experiments.At the same time,we improved the Hog feature detection in the traditional method.We can obtain the new Hog feature by decomposing the original image into multiple channels,and using weighted fusion to combine channel features.Finally,we compared the improved method with other methods to illustrate that the improved method is a bit.Secondly,there are only a few datasets of UAV,and the general datasets with rich targets and annotations are even more scarcer,which results in extremely imbalanced data between different classes of the datasets.For this problem,we make some relevant dataset expansion,so that the sample sizes are properly expanded.After the dataset expansion,we further eliminate the effect of the imbalanced data between different classes by adding a balance factor to the loss function.Thirdly,we improve the feature pyramid by using multi-level feature fusion.The target pictures taken by UAV often have large scale changes and are evenly distributed,and small and large targets in those pictures need to be detected.Therefore,the shallow and deep features of the images need to be combined.To this end,we mainly fuse the cross-layer features of the feature map,and then fuse the up-sampling part of the feature pyramid.Fourthly,we improve the up-sampling method.In the process of feature fusion,upsampling is a top priority and the basis for achieving feature fusion.Thus,we make comparative experiments on bilinear interpolation,deconvolution,pixel rearrangement,and mixed sampling methods.According to the experiment results,we replace the original bilinear interpolation method by the combination of deconvolution and pixel rearrangement.Finally,we introduce the attention mechanisms into the feature pyramid.By introducing a masking method in the feature pyramid and combining the basic SE structure,we introduce the attention mechanism into the seven-layer output feature map of the feature pyramid,and give different weights to each layer,and then implement the multi-feature fusion.