Vehicle Tracking Algorithm Based on Full Attention Mechanism

Vehicle tracking performance is significantly reduced due to problems such as blurring, target and background similarity, local occlusion, background blurring, dim light and so on. To overcome these problems, we propose a vehicle tracking algorithm based on full attention mechanism. Firstly, Swin Transformer based on attention mechanism is introduced to sufficiently excavate feature representation of target region and search region, and then use the Transformer encoder to fuse the above features. Finally, a simple stacked full convolution network is utilized to predict the target location. The experimental results show that the proposed algorithm has good robustness in blur scene, complex background, similar objects, dim light and scale transformation. It has better tracking performance than other popular trackers on the precision plots and success plots. Compared with the baseline STARK-S50, success score of our tracker is increased by 2.1% on the OTB2015 dataset, and the tracking speed can reach 47 fps.