Research On UAV Tracking Algorithm Based On Machine Vision

As an important branch in the field of computer vision,object tracking technology plays an important role in many social fields such as urban management,agricultural monitoring,rescue and disaster relief,and geological detection.Many excellent singleobject tracking algorithms have emerged,especially the siamese network-based tracking algorithm with its novel siamese branch structure has a good performance in the field of single-object tracking,so a series of follow-up algorithms are soon developed,but these algorithms still have some shortcomings when applied to UAV object tracking,such as in small objects,camera shake,fast motion,low resolution and other complex scenes with poor tracking results.The siamese network tracking algorithm using the anchor mechanism greatly increases the calculation of UAVs.To address the above problems,the work and research results of this thesis can be summarized as follows:1)The IMPSiam CAR algorithm is proposed to address the challenges of object tracking by UAVs,such as large scale change,easily obscured and similar object interference,the lack of calculation ability due to the limited computing equipment that UAVs could carry,which affects the tracking effect.CH-Res Net50 is used to extract features instead of the Res Net50 network in Siam CAR.A channel attention mechanism is introduced to make the model learn the semantic information of different channels and assign different weights to the channels according to the importance of the features,so that the algorithm pays more attention to the regions where objects exist.The experimental results show that the algorithm performs better for problems such as large object scale changes and occlusion,and the tracking accuracy and success rate are improved by 4.8%and 2.8% respectively on the UAV123 dataset compared with the benchmark algorithm Siam CAR,with a speed of 53.5 FPS,which could meet the requirements of UAV for algorithm performance and real-time performance.2)Aiming at the problems that UAV video objects have such as low resolution and camera shake,which lead to low accuracy,REFSiam CAR tracking algorithm which is built based on the framework of Siam CAR is proposed.Siam CAR is used as the initial tracker,and the CIOU border regression loss function is introduced to replace the IOU border regression loss function,and the aspect ratio,centroid distance and overlap area are also considered to solve the problem of non-overlap between the object frame and the prediction frame,so that the network converges quickly.The refined tracking module is used to further extract more detailed object features from the Siam CAR rough tracking results to obtain higher quality prediction frames and thus improve the tracking accuracy.The experimental results on the UAV123 dataset show that the success rate and tracking accuracy are improved by 3.4% and 2.3%,respectively,compared with Siam CAR,the performance is also better on the OTB2015 and VOT2018 datasets.REFSiam CAR can better deal with the problems of low resolution and camera shake.3)In order to solve the problem that small objects are poorly tracked by UAVs due to less effective information output from deep residual networks,a Siam MFF tracking method that introduces an efficient multi-scale feature fusion strategy is proposed.The method applies an efficient multi-scale feature fusion to aggregate features at different scales and constructs a multi-scale feature map containing more semantic information.Meanwhile,the ordinary convolution is replaced by deformable convolution to increase the convolutional operation field to enhance the feature extraction capability.The experimental results show that the proposed algorithm has improved the success rate and accuracy for small object tracking.