Multi-scale Target Detection System For Low-altitude Small UAV Based On Vision

With low-altitude airspace opening,the types and number of low-altitude slow-speed small targets have gradually increased,especially for multi-rotor small UAVs which have become main-stream aircrafts due to their simple operation and affordable price.Compared with other low and slow small targets,small UAVs endanger public safety and personal privacy more seriously because of their strong maneuverability and high concealment.Therefore,the researches on theory,tech-nology,and platform of anti-UAV systems have attracted great attention of governments at home and abroad,academia and industry.Small UAV is a typical low-slow-small target.The difficulties such as small detection area,irregular flight trajectory and complex low-altitude background lead to a huge challenge to detect it.This paper designs and implements a vision-based multi-scale low-altitude small UAV detection system.The system uses different detection methods for different airspace characteristics.One is the detection of weak targets in long-range airspace,and the other is the detection of small targets in near-ground airspace.Consequetly,the system implements full airspace coverage drone monitoring.The main research contents of this article include:(1)Aiming at the problem that weak target detection is susceptible to interference from com-plex ground information,a self-learning sky segmentation algorithm based on prior information is designed to extract the sky area.First of all,the inherent characteristics of the sky area under different weather,light and other conditions are analyzed in detail,and the prior information of the sky area is refined.Then,the image gradient map is obtained utilizing the Sobel operator.Un-der different gradient thresholds,different sky boundaries and their corresponding energy function values are obtained.By optimizing the energy function,the optimal initial sky area is obtained.K-means clustering is then performed on the initial sky area to obtain the true sky,and the other areas are segmented twice based on the true sky.Finally,the segmentation results are optimized through morphology operation.This algorithm only needs a single picture to achieve self-learning segmentation,and does not need to build a large number of data sets for training.In the experi-ments,this algorithm is compared with algorithms based on deep learning such as BiSeNet,and the results show that the algorithm is superior to other algorithms in detection effect and processing speed.(2)Based on the local visual saliency of weak and small targets,a spatio-temporal fusion spec-tral residual saliency detection method is proposed.This algorithm transforms the two-dimensional spatial information of the image into the frequency domain,avoids processing the image directly in space,and effectively retains the faint appearance characteristics of weak targets.At the same time,in order to suppress background noise,time-domain information is added to the original spectral residual significance model(Spectral Residual,SR).In order to simulate the decay of pre-sequence image information over time,time-domain recursive weighting is used to fuse the time-domain information.In the experiment,this algorithm is compared with 6 commonly used methods for de-tecting small and weak targets.The results show that the proposed algorithm can avoid the missed detection caused by the instability of long-distance imaging to a certain extent,and can also re-duce the false detection of high frequency noise in the background,with the detection performance significantly improved.(3)Based on the complex background of near-Earth airspace and the small target of drones,a small target detection network based on improved YOLOv3(You Only Look Once)is proposed.First,through the semi-automatic labeling method based on kernel correlation filter tracking,an-notate the target of the experimental video,generate a large number of training samples,and build a data set.Then,for the characteristc of small size,the proposed algorithm improves YOLOv3,combining finer features and context information to enhance the algorithm's ability to detect small targets(less than 500 pixels).Finally,the experiments are performed on test data sets and realistic scenes.The results show that the detection algorithm significantly improves the detection perfor-mance of small targets in complex backgrounds,and can achieve real-time detection in the system.(4)Design and implement the overall scheme of the system.This system can perform real-time monitoring of near-Earth airspace and long-range airspace simultaneously.The system uses a PTZ(Pan-Tilt-Zoom)camera to cruise near-Earth airspace,and uses a deep learning detection algorithm with high robustness and strong generalization performance to detect drone targets.In long-range airspace,a panoramic camera is used to monitor suspected weak targets(targets less than 100 pixels).First,a self-learning sky segmentation method based on prior information is used to segment the sky area to exclude the impact of the ground portion.Then,a saliency target detection algorithm is used to detect suspected targets in the sky,and then a detailed camera is used to zoom in and identify the target according to the target orientation provided by the panoramic picture.The system also includes the camera gimbal follow-up function and user interaction module.