PDRNN splits pedestrian tracking into modular RNNs to handle asynchronous sensor streams
A new recurrent neural network architecture handles asynchronous sensor data from IMUs and radio systems like 5G, estimating position, velocity, and orientation with separate ensemble models that avoid black-box error drift.
Pedestrian dead reckoning systems that fuse inertial and radio sensor data face a persistent problem: accelerometers, gyroscopes, and positioning receivers operate at different sampling rates and produce contradictory estimates during high-acceleration movement. A new preprint published on arXiv this week proposes PDRNN, a modular recurrent neural network architecture that treats each estimation task as an independent ensemble, allowing practitioners to swap components without retraining the entire pipeline.
The system separates orientation estimation (from gyroscope data), velocity or distance estimation (from accelerometer data), and optional absolute positioning (from synchronized 5G radio) into distinct RNN ensembles. Each ensemble outputs both a mean estimate and a variance estimate; a final fusion model weights these contributions using the variance signals to produce a combined pose estimate. Experiments on dynamic sports movement datasets show PDRNN outperforms both classical extended Kalman filter approaches and end-to-end black-box neural methods in accuracy and precision, while avoiding the error accumulation that plagues monolithic deep learning trackers. The modular design trades increased complexity—five or six separate models instead of one—for component-level control and interpretability.
Because each component is a recurrent network trained on time-series data, PDRNN can also forecast the next few timesteps of sensor readings along a reference trajectory, effectively filling gaps when radio transmissions drop or IMU data arrives late. The preprint does not include open weights or a public implementation, and the authors do not specify which 5G chipsets or IMU hardware were used in the sports trials. Practitioners interested in replicating the results will need to wait for a code release or contact the authors directly for dataset access and hyperparameter details. The next milestone will likely be a reference implementation and ablation studies isolating the contribution of each ensemble component to overall tracking accuracy.