Research On UAV Video Target Tracking Algorithm Based On Deep Learning

With the rapid development of modern technology,Unmanned Aerial Vehicle have been widely used in various fields,including military reconnaissance,material dispatch,power patrol,geological exploration and intelligent security.As one of the key technologies for UAV applications,target tracking technology has become a research hotspot in the field of machine vision.In UAV aerial video,tracking targets are often accompanied by background interference,scale changes,occlusion and other complexities,therefore,it is of great importance to study efficient and robust target tracking algorithms for UAV applications.This thesis investigates the UAV video target tracking algorithm based on deep learning,the main research work is as follows:（1）Aiming at the problem of low tracking accuracy of small targets in UAV video,an adaptive fusion network-based UAV target tracking algorithm is proposed.First,a deep network model is constructed based on the perceptual field module and Residual Network to effectively extract target features and enhance the effective perceptual field of the features.Then,a multi-scale adaptive fusion network is proposed that can adaptively fuse deep and shallow features of the network to enhance the expression of the features.Finally,the fused features are fed into the correlation filter model and the maximum confidence score of the response map is calculated to determine the tracking target location.The experimental results show that the algorithm achieves a high level of tracking success rate and accuracy,which effectively improves the performance of UAV target tracking algorithm.（2）Aiming at the UAV flight process,the target is prone to scale change,deformation and other problems.A scale-adaptive UAV target tracking algorithm based on Siamese network is proposed.First,combining the advantages of cross stage partial network and deeply separable convolution to construct a deep network model to extract target features,which optimizes the network structure and reduces the computational effort of the network model.Second,the extracted target template features are subjected to a correlation convolution operation with the search region features to obtain the feature response maps.Finally,a classification and regression network based on anchorfree is designed to calculate the position of the tracked target in the response map.The experimental results show that the algorithm can effectively reduce the impact of target scale change and deformation on the tracking performance.（3）In response to the problem that it is difficult to identify small targets for UAV multi-target tracking and difficult to match accurately when multiple targets are obscured and gathered,a UAV multi-target tracking algorithm based on full-scale apparent characteristics network is proposed.Firstly,a multi-scale target detector is constructed based on the YOLOv4 network model to enhance the detection capability of targets and provide data guarantee for subsequent tracking and matching.Then,a full-scale network model is designed to extract the target appearance features and enhance the recognition of similar objects.Simultaneously,the prediction of the motion state of the target using the Kalman filter.Finally,the target matching strategy based on the Hungarian algorithm is used to realize the matching association between detection results and tracking trajectories by combining the target appearance features and motion information,complete multi-target tracking tasks.The experimental results show that the algorithm effectively enhances the expression of features to small targets and has better tracking performance.