Optimizing Neural Inertial Classification: A Benchmark Study of Data-Driven Techniques

Inertial sensors are widely used for pedestrian activity recognition. Recent advances in deep learning techniques have significantly improved the inertial classification task's performance and robustness. However, a standardized benchmark for evaluating and comparing these methods remains lacking. Such a benchmark is critical for ensuring fair and consistent evaluation and future development. In this study, we aim to fill this gap by defining and analysing eleven data-driven techniques designed to enhance neural inertial classification networks. Our investigation focuses on three key components: network architecture, data augmentation, and data preprocessing. In addition, we conduct comparative analyses to identify the optimal window size for each dataset. This is a parameter that substantially affects model performance but is often overlooked. The experiments were conducted across seven datasets collected from 229 participants and with a total of 4482 minutes. Among the evaluated techniques, data augmentation through rotation and multi-head network architectures yielded the most consistent performance improvements. Our experimental results show that rotation-based augmentation and multi-head architectures consistently yield the highest gains, improving accuracy by up to 9.72% depending on the dataset and window length. We additionally quantify the effect of temporal window size, demonstrating that longer segments (2s) provide the largest average improvement, whereas shorter windows better suit real-time deployment. Finally, we propose a benchmarking strategy to support the future development and evaluation of deep learning models for inertial activity recognition.