Multi-template global re-detection based on Gumbel-Softmax in long-term visual tracking

In long-term visual tracking, target occlusion and out-of-view are common problems that lead to target loss. Adding a re-detection module to the short-term tracking algorithm is a general solution. However, the existing re-detection methods have limited accuracy, a large amount of calculation, and serious error accumulation, which seriously affect the algorithm’s long-term tracking ability. This paper proposes a flexible and accurate global re-detection module that enhances long-term tracking performance of the algorithm while improving re-detection speed. The proposed method innovatively uses three templates for global sampling to improve the re-detection accuracy. Then, Gumbel-Softmax is introduced into the re-detection module for accurate sampling, and a less number of target candidate boxes are output, which reduces the amount of computation. Finally, color feature is added to assist cosine similarity to locate the final target position more accurately. Four tracking algorithms are selected as benchmark algorithms (STMTrack, KeepTrack, SuperDiMP, and DiMP). The experimental results on five datasets (UAV123, UAV20L, LaSOT, VOT2018-LT, and VOT2020-LT) show that the long-term tracking ability of these algorithms can be effectively improved after adding the re-detection module. Especially on UAV20L, the accuracy and success rate of the improved STMTrack can be increased by 15.6% and 11.6% respectively.