To solve the problems of noise interference and error accumulation drift in low-cost MEMS IMU in pedestrian positioning applications, a tightly coupled inertial navigation method integrating Transformer and error-state Kalman filter (ESKF) is proposed. The central innovation of this approach is the synergistic integration of deep learning perception and classical physical constraints, alongside a novel adaptive observation noise adjustment mechanism driven by confidence weighting. Firstly, the multihead self-attention mechanism of Transformer is utilized to replace traditional recurrent neural networks to more effectively capture long-range temporal dependencies from raw IMU data, thereby achieving high-precision two-dimensional velocity regression. Secondly, a tightly coupled strategy is designed to map the Transformer network’s predicted velocity and its confidence into ESKF observation updates. Furthermore, a velocity error feedback mechanism is constructed using dual zero-velocity update (ZUPT) methods to correct heading drift and gyroscope bias. Experimental results based on the publicly available RoNIN dataset demonstrate that this method can effectively correct speed estimations, thereby enhancing overall trajectory accuracy. Compared with traditional RNN/LSTM algorithms, the proposed method exhibits superior robustness and the calculated trajectory matches the actual trajectory better. The 95th percentile of the absolute trajectory error and relative trajectory error were optimally reduced to 2.99 meters and 3.25 meters, respectively. Compared with the other two methods, the maximum positioning error has been reduced by the least 3.31 meters, representing a decrease of approximately 50.38%. The proposed method effectively improves the positioning accuracy of inertial pedestrian navigation, providing a certain reference for the autonomous positioning of consumer mobile devices in complex indoor environments.