Research Of UAV Target Detection And Tracking Algorithm Based On Deep Learning

Target detection and tracking has always been an important research direction in the field of computer vision.In recent years,it has been widely used in anti-terrorism war,public security monitoring,and regional inspections on the multi-rotor drone platform.With the development of artificial intelligence technology,target detection and tracking methods based on deep learning have made great progress in terms of accuracy and generalization ability.However,from the perspective of UAV,there are some problems such as small target size and complicated interference factors bring a great challenge to target detection and tracking tasks.Therefore,this paper mainly studies the UAV target detection and tracking algorithm based on deep learning.The specific work content is listed as follows:First of all,this article analyzes the key technical difficulties and problems for the target detection and tracking algorithm in the UAV usage scenario,and gives a general introduction to the basic theoretical knowledge involved in research methods.Aiming at the task of UAV target detection,this paper designs a UAV target detection algorithm based on improved YOLOv3.In order to enchance the algorithm’s detection performerce and target positioning accuracy,CIo U Loss function is introduced into the bounding box regression Loss,at the same time,Redefine the distance function in the anchor clustering algorithm,and improve the problem of excessive elimination of overlapping target boxes.What’s more,considering the characteristics of the drone video sequence,this paper also enrichs the training data by data augmentation and achieves advanced performance on different views and different scale targets.The improved M-YOLOv3 algorithm has reached a m AP of84.6% on this dataset.Aiming at the task of UAV target tracking,this study chooses Alex Net model as the feature extraction network on the basis of Siam RPN++ to ensures the real-time processing speed.The improved Siam＿Tracker reached 25 FPS on the airborne platform.In addition,due to the interferences such as scale variation,occlusion,deformation,and similar target,the problem of tracking drift occurs frequently,especially in the background clutters scenarios.In order to tackle this issue,this paper establishes a target motion estimation model based on the Kalman filter algorithm,so that the algorithm can combine with the target motion state to improve algorithm performance.Finally,a small multi-rotor UAV platform is finished to test the improved target detection and tracking algorithm,that is deployed to the NVIDIA TX2 airborne computing platform through network slimming and Tensor RT acceleration technology.The experimental results shows that the improved target detection and tracking algorithm satisfies the expected requirements.