A CNN and DRL-Based Heading Estimation Method for Pedestrian Dead Reckoning Localization

Pedestrian Dead Reckoning (PDR) has been widely applied in indoor localization due to its independence from additional infrastructure. However, its heading estimation is highly susceptible to environmental magnetic disturbances, leading to accuracy degradation. To address this issue, this paper proposes an adaptive heading estimation method based on Convolutional Neural Networks (CNN) and Deep Reinforcement Learning (DRL). First, the magnetometer signals are transformed into two-dimensional images using Continuous Wavelet Transform, and magnetic disturbances are identified through CNN-based classification. Then, a Deep Reinforcement Learning framework based on Deep Deterministic Policy Gradient is employed to dynamically adjust the fusion weights of the magnetometer and gyroscope headings, with a reward function that incorporates a heading smoothness constraint. Experimental results demonstrate that the proposed method achieves a magnetic disturbance classification accuracy of 99% and maintains a mean heading estimation error within 3°, outperforming benchmark algorithms. Furthermore, in real-world PDR localization experiments, the proposed adaptive heading estimation method achieves an average positioning error of 0.89 m, validating its effectiveness and robustness.