NanoMST: A Hardware-Aware Multiscale Transformer Network for TinyML-Based Real-Time Inertial Motion Tracking

Deep learning-based inertial navigation remains a formidable challenge due to the intricate temporal dynamics of human motion and the stringent computational constraints of edge devices. This study introduces NanoMST, a highly efficient multiscale transformer architecture designed for precise pedestrian inertial motion tracking with minimal computational overhead. The proposed model integrates a hierarchical multiscale embedding strategy with a scale-adaptive attention mechanism, effectively capturing motion patterns across diverse temporal resolutions while optimizing efficiency through hardware-aware quantization. With 298K parameters and 7.59M floating-point operations, Nano Multiscale Transformer (NanoMST) achieves performance comparable to substantially larger models while maintaining an exceptionally low computational burden. Extensive evaluations on benchmark datasets, including OxIOD, robust neural inertial navigation (RoNIN), and RIDI, yield average trajectory errors of 2.68m on RoNIN, 1.64m on RIDI, and 1.80m on OxIOD. The quantized 8-bit implementation reduces the model size from 1.23MB to 0.41MB while retaining 94% of the original model’s accuracy. Profiling on edge devices confirms real-time feasibility, with inference latencies ranging from 0.18 to 0.96 ms across various smartphone generations and an average throughput exceeding 6000 samples per second, surpassing contemporary architectures, such as IMUNet and CTIN. This study illustrates an efficient engineering approach for deep learning-based inertial tracking, demonstrating that high-precision sequential motion estimation can be achieved with a minimal computational footprint. The efficiency and real-time capability of NanoMST make it particularly suitable for deployment in resource-constrained environments, including mobile, wearable, and Internet of Things applications.