The Research And Application Of UAV Detection

With the progress of the integrated control technology of hardware,the using threshold of civilian UAVs is more and more low,application scope is more and more widely.This brings a lot of fun and convenience to the public life,but the illegal use of UAVs makes the regulation of navigation safety more and more difficult.The counter for UAVs and detection becomes more important.In practical UAV detection applications,UAV targets are usually small and the environment is complex,so the environmental robustness and detection accuracy of the detection algorithm,especially the accuracy of small target detection,are highly required.To solve this problem,a random splicing data augmentation algorithm based on tiled image pyramid and an adaptive feature fusion module combined with attention mechanism are proposed to improve the detection accuracy.In order to improve the robustness of the algorithm in UAV detection task,UAV dataset is made for possible scenarios,and experiments are carried out on this dataset to select the algorithm structure most suitable for the task in this paper from numerous candidate algorithm versions.The main contributions of this paper are as follows:1)A random splicing data augmentation algorithm based on tiled image pyramid(TIP)is proposed to effectively expand and balance the target distribution of the dataset and improve the detection accuracy,especially the small object detection accuracy.Based on the basic idea of image pyramid scaling to generate images of different scales,the algorithm splicing all scale images into one image to achieve the expansion of the number of targets,improve the accuracy of object detection,and avoid the phenomenon of over-fitting through the completely random selection of images and splicing positions.2)An adaptive feature fusion module(AWFM)combined with attention mechanism is proposed to optimize the utilization of multi-scale feature information,and balance the detection accuracy and running speed of the algorithm.This AWFM module filters the importance of features in the channel domain and spatial domain to obtain the weights of feature fusion by adaptive training.Module uses these weights to control features contribution of different level fusion,thus,the feature information of each scale can be better integrated and utilized to improve the detection accuracy.3)Application and verification of UAV detection scheme.According to the requirements of the UAV detection task for the algorithm robustness of complex scenes,a UAV dataset meeting the requirements is made,and experiments are carried out on this dataset.Through comparison and verification,an object detection algorithm,with background robustness and balance between real-time performance and detection accuracy,is selected.The detection accuracy of the proposed algorithm can reach about 94%in the UAV dataset of complex scenes,and the accuracy of small target detection can reach about 26%,the minimum target can be detected at about 20*20 pixels.