Research On Low-Altitude Drone Detection Method Based On Improved YOLOv5

Due to the drone industry’s quick growth,low-altitude research has focused heavily on drone inspection technologies.Traditional detection methods,however,find it challenging to extract useful target feature information since the low altitude background is complicated and unpredictable,the number of drone target features is limited,and the number of small targets is significant when detected at long distances.To solve the aforementioned issues,the improved YOLOv5-based low-altitude drone detection approach is investigated in this paper to enhance the detection efficiency of small low-altitude drone targets.The paper’s primary research components are as follows:To solve the issue of limited data volume in the area of low-altitude drone target detection,the particular low-altitude drone target dataset is created in this paper.Firstly,the original dataset is constructed by field photography,crawler acquisition,self-collection,and borrowing from existing datasets.Secondly,to successfully handle the effects of complex environmental factors like light and weather,the true outdoor environment of drone flight shooting is reproduced by utilizing traditional data augmentation algorithms on the original data set.Then,to increase the fraction of small targets in the dataset and enrich the detection background,the proposed random stitching data augmentation,intercepted embedding data augmentation,and large-scale jittering data augmentation approaches are applied.Finally,the obtained low-altitude drone target dataset,which includes 21,844 sample photos in total,provides a solid dataset basis for the ensuing research.To address the issue that low-altitude drone targets have a small target share and low pixel count in a straightforward background with little interference information and clearances.The small-scale drone detection technique based on bi-directional feature fusion is suggested in this paper.Firstly,S-E Attention,a serial attention module,is built to improve the model’s global sensing capability while capturing the local relationship of features.This is done to make up for the information loss that occurs when shallow feature channels are switched over to deep channels.Secondly,Content-Aware Re Assembly of FEatures upsampling,which uses the surrounding data in a larger perceptual field for modeling,is used to replace the original model’s high-level nearest-neighbor interpolation upsampling.Finally,to reduce the loss of low-level information during the downsampling process,the low-level downsampling mechanism of the original model is modified to a context-aware downsampling module.The pooling features and dilated convolution features are stacked throughout the downsampling process.Numerous sets of experiments are used to confirm the performance of the small-scale drone identification technique based on bi-directional feature fusion.To address the issue of low-resolution and undetectable feature information of low-altitude drone targets in complex backgrounds such as buildings,cities,woods,mountains,and light changes.The feature-enhanced drone target detection algorithm based on feature enhancement is proposed in this paper for complex scenes.Firstly,the feature extraction network C3 module and Coordinate Attention location attention are fused to combine multi-scale local feature information and global feature information.This will further improve the network’s sensitivity to information such as direction and location.Secondly,the Backbone Enhancement Module,a backbone network enhancement module is built by novel Cross convolution and channel attention mechanisms.The module is intended to investigate horizontal and vertical gradient information to concentrate on deeper information mining and resolve the issue of limited features extracted by conventional convolution.Finally,to reduce detail information loss and aggregate multi-scale information,the C3 ASPP module is designed to replace the Spatial Pyramid Pooling-Fast module of the backbone network.Multiple sets of experiments are used to validate the advanced feature enhancement-based drone target recognition method in complex scenes.