SpikePR: Position Regression With Deep Spiking Neural Network

Abstract

Data-driven human localization technology has been on the rise with advancements in end-to-end artificial neural networks (ANNs) in recent years. Different from the traditional pedestrian dead reckoning (PDR) algorithms, the data-driven method can significantly reduce cumulative error over time arising from integration and improve the accuracy and efficiency of localization. However, the computation complexity of ANNs imposes high requirements on hardware conditions and heavily hinders its application on mobile devices. Targeting the above challenges, we design a Position Regression algorithm with a deep spiking neural network (SNN, called SpikePR)—an architecture inspired by biological neurons—to regress the user’s position when collecting a sequence of raw inertial measurement unit (IMU) measurements from mobile devices. This architecture integrates ANNs and SNNs with a leaky integrate-and-fire (LIF) mechanism due to its low-power computation with binary spikes and capability to model the temporal dynamics in time-series data. We conduct extensive experiments on four open-source datasets with the proposed SpikePR algorithm. The experiment results demonstrate that compared to the state-of-the-art position regression algorithms, the proposed SpikePR can save more than 90% energy consumption while achieving similar location errors.