Research On UAV Detection And Tracking Algorithm Based On Deep Learning

Target detection and tracking is an important research direction in the field of computer vision,and it plays a key role in many application fields such as intelligent transportation,unmanned driving,autonomous robot navigation,disaster or injury detection.Target detection algorithms based on deep learning technology are usually trained on the existing publicly available large-scale general-purpose data sets.Since these data sets are mostly for sparse targets in natural scenarios,the number of small target samples is small,making these excellent algorithms effective effect of small-scale target detection is not very good.This topic is the target detection and tracking in the UAV application scenario.Due to many problems such as target deformation,complex changes in the scene,and small target size during the tracking process,the target detection from the perspective of the UAV and tracking is greatly challenged.In this paper,the existing mainstream deep learning-based target detection and tracking algorithms are researched,analyzed and further improved to improve the accuracy of small target detection and tracking in UAV application scenarios.The main research of this paper includes the following three aspects:(1)This paper optimizes the existing YOLOv3 algorithm from the three perspectives of data expansion,network structure design,and loss function,and solves the problem of high missed detection rate of small target detection from the perspective of drones.(2)Although detection algorithms based on deep learning are much better than traditional detection algorithms based on hand-designed feature extraction operators,deep learning algorithms usually use a large number of convolutional layers or fully connected layers,which makes the algorithm have a huge amount of parameters,The huge amount of parameters makes the network’s forward reasoning very time-consuming,making it almost impossible to deploy and detect on the airborne end of the drone.Therefore,according to the common model compression method now,the improved YOLOv3 network is pruned.Operation to speed up the detection speed on the airborne end of the UAV.(3)In response to the challenges encountered in the actual tracking process,the existing deep learning-based target tracking algorithm Siam RPN is studied and improved.In order to solve a series of challenges faced by the target in the tracking process and improve the robustness of the tracker,first,the Siam RPN algorithm is used to extract the features of the 5-layer main network to replace,and the Res Net unit is used to build a deeper network structure to make the extracted features are more abundant and comprehensive,and can cope with more complex environments.Secondly,the matching template information is enhanced.The original algorithm only relies on the first frame of target image as a template image,so it is not enough to deal with complex targets and changes in the environment.The improved algorithm fuses the first frame template information of the video sequence,the feature information extracted from the previous frame of the current frame,and the historical fusion information and uses it as the tracking template of the current frame.In addition,due to the complex changes in the background during the tracking process,the newly fused template information is noisy,so a proportional weighting is made between the template information in the first frame and the newly fused template information,making the template information more robust.