Research On Target Tracking Algorithm Based On UAV Vision

UAVs equipped with high-quality shooting equipment are widely used in smart transportation,logistics transportation,filming,military investigation,disaster monitoring,etc.due to their advantages such as lightness,flexibility,and good concealment.Visual target tracking technology is one of the key technologies for drone applications,which can greatly enhance the autonomy and intelligence of drones.The rapid development of computer vision technology has superior performance in processing various visual tasks.However,due to the various challenges of video shot by drones,such as occlusion,deformation,scale change,and background interference,traditional algorithms are not suitable.For processing videos and images taken by drones.Therefore,the research of visual target tracking algorithms that are accurate,efficient and suitable for drone video is of great significance for the application of drones.The thesis studies the algorithm of target tracking technology based on UAV vision.The main research work is as follows:(1)Aiming at the situation that the altitude changes constantly during the drone flight and the target has serious scale changes,a multi-scale fast UAV target tracking algorithm based on gcForest is proposed.Using gcForest decision tree forest network instead of traditional convolutional network to extract features effectively reduces the complexity of the network and improves the efficiency of the algorithm.Then,the multi-scale change method is used to enrich the target sample information,and three scale input channels are added to the input layer of the network to adapt to the input of the multi-scale image,to overcome the scale problem and improve the accuracy of the network’s representation of target features.Finally,the use of compressed sensing theory for feature dimensionality reduction further improves the calculation efficiency and reduces the feature redundancy.Experimental results show that the algorithm can achieve stable tracking in various common scenarios of UAVs,and can effectively improve the tracking rate,especially in the scene of scale changes.(2)Aiming at the situation that the UAV video background is complex and there are similar objects around the tracking target,an Octave convolution UAV anti-target interference tracking algorithm is proposed.A loss function of spatio-temporal correlation is designed,which fully considers the target position information of the previous frame,and reduces the influence of similar objects on position judgment through distance penalty.Then use Octave convolution instead of traditional convolution to improve network computing efficiency and optimize memory footprint,providing a more effective model for the tracking algorithm.Experimental results show that the algorithm can track targets stably under various UAV actual flight scenarios and effectively reduce the impact of similar objects on tracking performance.(3)Aiming at the problem that the target in the UAV video is prone to morphological change and easy to be blocked,a tracking algorithm for UAV target in the occlusion and deformation scenarios is proposed.The algorithm designed a 5-order Siamese network.The dual-channel network extracts template features and current frame features.To improve the accuracy of the target template,an adaptive network model is constructed combining Siamese network features and adaptive strategies.Then,according to the constructed template image set,the predicted target of each frame is processed and optimized,and the regression model is introduced to improve the prediction accuracy.Experimental results show that the algorithm achieves better tracking accuracy in different scenarios and improves tracking accuracy in occlusion and deformation scenarios.