The invention discloses an optimal label selection method for an RFID equipment-free human body tracking system, which comprises the following steps: S1, region division: dividing a monitoring region into N positions; s2, feature extraction and calculation: extracting a mean value and a variance of T RSSIs sampled at each position within a period of time, and establishing a mapping relationship between the distribution of the RSSIs in the period of time and the positions of the RSSIs; s3, constructing a deep learning model, and analyzing a corresponding position sequence according to the RSSI sequence of the T, namely a real movement track of a human body; and S4, label layout mode selection: preferentially selecting labels according to the classification accuracy of the deep learning model for the positions. Through the deep learning model, while the human body tracking precision is maintained, the number of labels is reduced, the flexibility of the model is improved, precision reduction is avoided, the performance of processing a long-path sequence is improved, and the gap between training and inference of a position sequence prediction task is filled up.