Single Target Tracking for UAV Based on Deep Learning
Unmanned aerial vehicle(UAV)single target tracking refers to real-time processing of videos captured during UAV movement,accurately and stably tracking a moving target.UAV single target tracking is greatly affected by the environment,with issues such as changes in lighting,background interference,target occlusion,and interference from similar targets,resulting in the need for further improvement in tracking accuracy.We focus on these issues and creatively optimize the SiamRPN++ model and loss function.The main research contributions are as follows:in terms of network backbone,we introduce the attention mechanism network structure SENet and combine it with the original ResNet50 model to form Se_ResNet50,improving the accuracy and effectiveness of singles target tracking.In terms of loss function,we use Balanced L1 Loss to enhance the key regression gradients and achieve a more balanced training in classification,overall localization,and precise localization.Based on the structure of SiamRPN++,we optimize the Backbone and Loss functions.Experiments were conducted using the ILSVRC2013 and ILSVRC2014 DET datasets for training and VOT2018 and OTB100 as test datasets to verify training accuracy.Ultimately,tracking accuracy was improved to a certain extent compared to the original.
国家科技期刊平台
NSTL
万方数据
维普
