On Neural Inertial Classification Networks for Pedestrian Activity Recognition

TL;DR

This work addresses the lack of standardized benchmarks for neural inertial classification in pedestrian activity recognition by evaluating ten data-driven techniques across architecture, augmentation, and preprocessing. Using four real-world HAR datasets (936 minutes from 78 participants), the study demonstrates that rotation-based data augmentation and multi-head network designs yield the largest, most consistent gains, even when baselines are already high. The findings offer practical benchmarking strategies and guidance for deploying inertial classification networks in real-world pedestrian navigation and related applications. Overall, the paper provides a comprehensive, cross-dataset assessment of methods to improve neural inertial classification, highlighting when and why certain techniques are advantageous.

Abstract

Inertial sensors are crucial for recognizing pedestrian activity. Recent advances in deep learning have greatly improved inertial sensing performance and robustness. Different domains and platforms use deep-learning techniques to enhance network performance, but there is no common benchmark. The latter is crucial for fair comparison and evaluation within a standardized framework. The aim of this paper is to fill this gap by defining and analyzing ten data-driven techniques for improving neural inertial classification networks. In order to accomplish this, we focused on three aspects of neural networks: network architecture, data augmentation, and data preprocessing. The experiments were conducted across four datasets collected from 78 participants. In total, over 936 minutes of inertial data sampled between 50-200Hz were analyzed. Data augmentation through rotation and multi-head architecture consistently yields the most significant improvements. Additionally, this study outlines benchmarking strategies for enhancing neural inertial classification networks.

Figures (12)

• Figure 1: Baseline network architecture. The output differs between each task and dataset.
• Figure 2: Head2 architecture. One head receives the accelerometer readings and the other receives the gyrosopce readings.
• Figure 3: Head3 architecture. Each head receives the accelerometer and gyroscope readings along the x,y, and z axes.
• Figure 4: Amount of time for each of the four classes in the RIDI dataset.
• Figure 5: Amount of time for each of the six classes in the MotionSense dataset.