Dynamic target tracking method for optical flow estimation in unmanned aerial vehicle perspective
Unmanned Aerial Vehicle(UAV)dynamic target tracking is one of the critical technologies in the search and rescue field and environmental monitoring,etc.How to accurately discriminate and capture the motion information of the moving targets is of critical importance.Currently,most trackers focus on modeling the appearance features of the targets without considering their motion information.Learning-based optical flow estimation methods estimate the motion information of moving target displacement while describing the appearance features.Therefore,we combine Transformer with optical flow estimation to propose a compact feature encoder that establishes correlation between global contextual information and reinforces regional features through a self-attention mechanism.Moreover,a cross decoder is proposed to perform flow regression,which combines iterative refinement and cross-attention mechanism to remove a large number of refinement steps and improve the processing speed.Through the up-sampler,the optical flow estimation is achieved with high feature resolution from UAV viewpoint.Our experimental results show our method achieves higher accuracy and robustness compared with those of other advanced methods.
moving target trackingoptical flow estimationattention mechanismmotion information
国家科技期刊平台
NSTL
万方数据
