Small UAV Target Detection Under Dynamic Scenes

In recent years,the rapid development of UAVs(Unmanned Aerial Vehicles)has posed a serious threat to public security,personal privacy and military security,etc.The number of dangerous accidents caused by UAVs is growing rapidly.Thus,discovering unknown UAVs fast and reliably becomes more and more important.Among existing UAV detection techniques,the vision-based approach is cost-effective and easiest to implement.Therefore,the study of vision-based UAV detection technology is of great importance to prevent the UAV threat.However,this task also faces many challenges,for example,the detection background is complex,the target is small,and the flight speed is high.Considering the low detection accuracy problem caused by these factors,this research mainly studies the small UAV detection technology in dynamic scenarios,including the following two aspects:Firstly,this thesis proposes a small UAV target detection method based on UAV motion information in videos.This method focuses on use the motion information of the UAV in the video to detect the UAV under dynamic scenes.First of all,considering that it is difficult to excavate the UAV motion information due to the camera motion,a video stabilization method is employed to eliminate the motion of the camera.Then,the lowrank analysis method is used to detect the UAV proposals.Finally,the thesis employs the CNN(Convolutional Neural Network)to extract the objects’ feature and input it into SVM(Support Vector Machine)to further classify and confirm the true target of UAV.The experimental results show that the proposed algorithm has a higher accuracy than the existing detection methods.Additionally,this method also has some shortcomings.For example,the video stabilization algorithm can’t completely eliminate the camera motion,and the SVM could easily classify the small target as a false target.In addition,this method is very time-consuming and is hard to achieve the real-time UAV detection.Secondly,to overcome the shortage of the first task,this thesis proposes a small UAV detection method based on a single image.Inspired by the regional convolution neural network in object detection,this thesis designs a feature fusion network and integrates it into the framework of the Faster R-CNN network.Unlike the Faster R-CNN which uses the last layer feature,the proposed feature fusion network fuses the features of each network layer.The fused feature combines the advantages of high-level and low-level features,so it can detect multi-scale targets well.In order to filter out the false target of UAV,the algorithm finally uses the nearest neighbor distance matching method to track the UAV.Experiments show that the proposed algorithm can effectively improve the detection accuracy with less computation time when compared with the first task’s method,and the other state-of-art methods such as Faster R-CNN,SSD.