Pedestrian Dead Reckoning using Data-Driven and Physics-Informed Machine Learning

Abstract

Many applications such as indoor navigation, search and rescue operations, and surveying, require accurate localization in challenging environments. In environments where absolute positioning (using for example GNSS receivers) fails due to dense obstruction of the required signals, positioning algorithms rely solely on dead reckoning, often using measurements from an Inertial Measurement Unit (IMU). This dead reckoning process suffers from integration drift due to the accumulation of errors in the sensor measurements. Pedestrian Dead-Reckoning (PDR) algorithms can reduce this integration drift in cases where the IMU is held by a walking person, by leveraging patterns from the periodic walking motion. This thesis investigates a state-of-the-art PDR algorithm from Liu et al. that combines an Extended Kalman Filter (EKF) with a velocity-predicting neural network that corrects the filter in the measurement update to mitigate integration drift. The focus is on finding if its performance can be improved by adapting the neural network in the algorithm. First, adaptations of the network’s parameters have been experimented with to investigate their effect on the algorithm’s accuracy and search for a bottleneck that limits it. This bottleneck was found to be a bias in the predicted velocity by the network, as this violates the EKF assumption that the error in the measurement update is distributed with zero mean. The second part investigates how using ideas from the Physics-Informed Neural Network (PINN) as an alternative to the data-driven neural network in the PDR algorithm affects its performance. Four network architectures have been trained using a loss function that includes a penalty for errors in the physics of the predicted velocities of the system. Training these PINN-inspired networks required a much longer training time. The results show that using this physics loss does not show significant improvements in the accuracy of the PDR algorithm, as the bias in the velocity prediction is not addressed by the physics loss. This thesis concludes that the main limitation of the neural network in the PDR algorithm by Liu et al. is a bias in the predicted velocity, and that a PINN is unable to provide significant improvements in the accuracy of the algorithm, whilst costing much more computational resources to train. It is therefore not recommended to use a PINN in this PDR algorithm, and the optimal configuration of the network was found to be a slightly adapted version of the network used by Liu et al., using IMU and device tilt data sampled at 50 Hz as network input. Future research could focus on understanding the origin of the bias in velocity predictions and to mitigate it once its source is known.