For wearable device motion monitoring, the accuracy rate is not high, and the background is complex. The motion tracking problem can be studied deeply using deep learning methods. In this paper, we design an instance-aware feature extraction network SiamRPN++, which can effectively distinguish different instances of the same object and avoid the “drift” of the tracker due to similar interfering objects. Secondly, a template update strategy based on mapped gradient (PGD) is proposed to address the problem of dramatic changes in target morphology during the motion tracking process. Finally, the PGD-based algorithm updates the templates effectively during the tracking process to avoid the failure of the tracking task due to the change in target morphology. The experimental results demonstrate that the target tracking algorithm using SiamRPN++ improves by a 12.24% success rate and 14.41% accuracy when tested on the target tracking dataset. The target tracking algorithm using the PGD template update strategy improves by 9.67% in success rate and 14.61% in accuracy. Wearable devices based on deep learning algorithms can achieve scientific guidance of sports exercise to provide strong support for avoiding the risk of high-intensity exercise and escorting safe and healthy sports.