Abstract: In mine accident rescue operations, the underground environment becomes extremely complex. Obstructions such as tunnels and rocks, coupled with changes in physical space and the presence of high-concentration gases, lead to attenuation, reflection, refraction, and the formation of multiple propagation paths of ultra-wide band signals. These factors make Non-Line-of-Sight (NLOS) and multipath effects very pronounced, which severely affects positioning accuracy. To address these issues, a novel algorithm based on a prior-data-driven Gaussian mixture model (GMM) and acceleration-collaborative adaptive extended Kalman filter (PGA−AEKF) is proposed. The algorithm first uses the Akaike information criterion and the Bayesian information criterion to determine the number of Gaussian mixture distributions. It then employs particle swarm optimization and simulated annealing algorithms to optimize the expectation maximization algorithm to determine the parameters of each distribution. The improved GMM is used to fit the collected time-delay prior data distribution. Subsequently, the real-time NLOS probability value of the scene is obtained using geometric methods, and an adaptive factor is generated in combination with the results of abnormal acceleration judgments of pedestrian or vehicle motion. Finally, the adaptive factor is used to adjust the size of the noise covariance matrix to improve positioning accuracy. To verify the performance of the algorithm, trajectory tracking experiments were conducted in Yunlongshan Tunnel and underground space smoke environment. The results show that in the tunnel scenario, the proposed PGA−AEKF algorithm achieved a significant reduction in average error compared to pure inertial measurement unit algorithm, least squares method, and extended Kalman filter algorithm, with reductions of approximately 86.67%, 69.23%, and 38.46%, respectively. In the smoke scenario, the average error was also significantly reduced by approximately 83.78%, 60.00%, and 25.00% compared to the aforementioned algorithms. The study shows that the PGA−AEKF algorithm can effectively deal with NLOS interference, suppress noise and errors.