Tiny_ML_on_HAR_Dataset

This Project was based on TinyML-Based Lightweight Deep Learning Model for Human Activity Recognition on Edge Devices which also resulted in an conference paper presented at WCAIAA 2026 conference held at NFSU GOA by SCRS

TinyML-Based Lightweight Deep Learning Model for Human Activity Recognition on Edge Devices 📱⚡

Overview

This repository contains the implementation of our research paper “TinyML-Based Lightweight Deep Learning Model for Human Activity Recognition on Edge Devices”, presented at a conference. The project explores how compact deep learning architectures can be optimized for real-time human activity recognition (HAR) using smartphone sensor data, while being deployable on resource-constrained edge devices.

Key Highlights

• 📊 Dataset: UCI Human Activity Recognition dataset (accelerometer + gyroscope signals).
• 🧹 Preprocessing: Noise filtering, normalization, segmentation into overlapping time windows.
• 🧠 Models Implemented:
  • NormalCNN1D – baseline convolutional model.
  • MobileNet1D – lightweight depthwise-separable CNN, best performing (92.77% accuracy).
  • CNN + BiLSTM Hybrid – combines spatial and temporal feature extraction.
• ⚙️ Optimization Techniques:
  • Weight pruning (60% reduction in parameters).
  • Post-training INT8 quantization (model size reduced to ~1.4 MB).
  • Knowledge distillation for lightweight student models.
• 🚀 Deployment: Models converted to TensorFlow Lite and tested on microcontroller-class hardware (Arduino/ESP32) for real-time inference.

Results

• MobileNet1D achieved 92.77% accuracy with balanced precision-recall trade-off.
• Quantized MobileNet1D retained nearly identical accuracy while reducing size to under 1.5 MB.
• Real-time inference achieved with latency <100ms and power consumption <50mW on edge devices.

Applications

• Health monitoring & fitness tracking.
• Fall detection & assistive technologies.
• Gesture control in IoT and wearable devices.

Future Scope

• Real-world latency and energy validation on embedded hardware.
• Exploring transformer-based TinyML architectures.
• Integration of federated learning for personalized HAR.

Presented at:

7th World Conference on Artificial Intelligence: Advances and Applications (WCAIAA 2026)

• Organized by: National Forensic Sciences University (NFSU), Goa, India
• Technically Sponsored by: Soft Computing Research Society (SCRS)
• 📅 January 30–31, 2026

🏅 Conference Presentation Certificate

🎤 Presented this work at WCAIAA 2026

🔗 View Certificate:
https://drive.google.com/file/d/1ci1EG9D85K83cUhpkDl4lZQx8boihCI-/view?usp=sharing

---