A device-free human tracking method based on selected tags was proposed to address the problems of existing human tracking systems, such as large tracking errors, large changes to the environment caused by hardware deployment, and inconvenience for users to wear the device. The radio frequency (RF) power signals of the human body at different positions were collected, and the relationship between the positions of the human body and the power signal was analyzed. In order to mitigate the effect of signal instability on the system, the features of the RF power signal were extracted as the dataset for model training. Different deep learning models were established by using the features of the RF power signal, and the tracking results and label usage under different training methods were compared to select the appropriate model structure and training method. The general opinion and the specific steps of the optimal tag selection method were given after analyzing the experimental results. Experimental results show that the location system based on the optimal tag selection method is feasible. The tracking error was 0.19 m, which was about 5 cm lower than that of the WallSense system, and the number of tags was reduced by 66.7%. The impact of tag deployment on the environment was obviously reduced, and the tag redundancy was reduced while the accuracy was ensured.