Multi‐template temporal information fusion for Siamese object tracking

The object tracking algorithm based on Siamese network often extracts the deep feature of target to be tracked from first frame video sequence as a template, and uses template for whole process. Because manually annotated in is more accurate, these algorithms have stable performance. However, it difficult adapt changing features only using extracted frame. Inspired by fusion transformer, this paper proposes update module called multi-template temporary information (MTFM), which can trained offline. By fusing multiple time series, always changes appearance In order train MTFM, training method series data Mean Square Error (MSE) loss function. This MTFM SiamFC++ tracker, obtains good experimental results three challenging datasets, including VOT2016, OTB100 GOT-10k. running speed graphics processing unit (GPU) maintained at about 200fps, exhibits real-time