Edgeefficient human activity recognition using a quantized patchbased transformer

Area of Science:

• Computer Science
• Artificial Intelligence
• Edge Computing

Background:

• Deep learning for Human Activity Recognition (HAR) faces deployment challenges on edge devices due to resource constraints.
• Inference latency, memory footprint, and computational capacity are key limitations for edge HAR systems.

Purpose of the Study:

• To propose a lightweight, patch-based Transformer architecture for efficient HAR in resource-constrained edge environments.
• To evaluate the proposed model's performance using architectural variations and quantization strategies with TensorFlow Lite.

Main Methods:

• Developed a compact Transformer model with patch embedding and multi-head self-attention for HAR.
• Conducted experiments varying patch lengths and encoder depths on WISDM and PAMAP2 datasets.
• Applied quantization techniques (FP32, FP16, INT8 dynamic, INT8 full) using TensorFlow Lite for optimization.

Main Results:

• The baseline model achieved high accuracy (96.77% on WISDM, 98.48% on PAMAP2).
• TFLite-INT8-Dynamic quantization reduced model size by up to 95.55% and improved latency by ~99% with minimal accuracy degradation.
• Moderate temporal patching balanced representation learning and computational efficiency.

Conclusions:

• The proposed patch-based Transformer model offers an effective balance between HAR performance and computational efficiency.
• The model demonstrates strong deployment feasibility for resource-constrained edge environments via TensorFlow Lite.
• This work provides a scalable solution for edge intelligence applications in HAR.