Deployment-Oriented Evaluation of Lightweight IMU-Based Human Activity Recognition: On-Device Efficiency and Streaming Feasibility

초록

On-device Human Activity Recognition (HAR) requires balancing accuracy and deployment efficiency on constrained hardware. We present Lightweight Human Activity Recognition (L-HAR), a controlled comparison of Baseline and Lightweight Temporal Convolutional Network (TCN), Transformer, and Long Short-Term Memory (LSTM) models using data sampled at (Formula presented.) from three inertial measurement units (IMUs) worn by one participant. Evaluations covered classification, model footprint, multiply–accumulate operations (MACs), offline CPU latency, Desktop/Raspberry Pi streaming, 8-bit integer (INT8) quantization, and software-estimated Raspberry Pi energy efficiency. Lightweighting reduced parameters by up to (Formula presented.), model size by (Formula presented.), and MACs by (Formula presented.) – (Formula presented.), with accuracy reductions of (Formula presented.) – (Formula presented.) percentage points (pp) and sub-millisecond offline latency for all Lightweight models. Baseline and Lightweight models sustained approximately (Formula presented.) message/inference rates on Desktop, whereas no Raspberry Pi configuration reached (Formula presented.) inference. Among Lightweight models, Transformer achieved (Formula presented.) End-to-End (E2E) latency and (Formula presented.) inference. INT8 quantization changed accuracy and F1-score by less than (Formula presented.) pp; Quantized TCN achieved the best Raspberry Pi streaming result ((Formula presented.) E2E; (Formula presented.)) with an estimated (Formula presented.) per inference. These results demonstrate architecture- and backend-dependent deployment behavior and the need for direct target-platform evaluation. © 2026 by the authors.

키워드