Tech Stack

This page proposes the technology stack and frameworks used for the FMCW radar project

Model Training

Layer	Component / Tool	Purpose / Functionality	Language / Framework
🛰️ Data Acquisition	TI mmWave Radar (IWR6843 / IWRL6432)	Captures FMCW radar signals (range–Doppler–angle) for human gait signatures	TI Firmware / C
	mmWave SDK / Demo Visualizer	TI tool for configuring radar profiles and exporting raw data	TI SDK / C / UART
	UART Manager	Streams radar frames to PC for logging and preprocessing	Python / C++ (serial)
⚙️ Preprocessing & Feature Extraction	Radar Signal Processor	Parses ADC/point cloud frames, applies range–Doppler FFTs	Python (NumPy, SciPy)
	Micro-Doppler Feature Generator	Converts radar cubes into spectrograms or time–frequency maps	Python (NumPy, Matplotlib)
	Data Labeling / Metadata Management	Annotates participant sessions, adds YAML/JSON metadata	Python
💾 Dataset Management	Git + GitHub Repo (spe-uob/2025-FMCWRadar)	Stores code, configs, and versioned dataset metadata	Git / GitHub
🧠 Model Training & Evaluation	PyTorch	CNN model training and inference on ?	Python / PyTorch
	Hydra	Config management for experiments and hyperparameter sweeps	Python (hydra-core)
	TensorBoard	Visualize training metrics, losses, and comparisons	TensorBoard
	ONNX + ONNX Runtime	Model export and post-training quantization	Python (onnxruntime)
⚡ Deployment & Runtime	?	No idea	?

Radar Data Visualizer

Layer	Technology	Purpose
Sensor Layer	TI IWR6843 / IWRL6432	Produces real-time radar frames
UART Parsing Layer	Python + PySerial + threading	Reads binary frames and emits structured radar data
Frame Parsing	Custom parser (parse_frame.py) + NumPy	Converts binary stream → range, doppler, point cloud
Backend Server	FastAPI + WebSocket + PyTorch/ONNX Runtime + Hydra	Receives parsed data, runs model inference (if needed), streams to frontend
Frontend	React.js + Three.js + TailwindCSS / ShadcnUI	Visualizes live point clouds, classification labels, and confidence
Communication	WebSocket (ws://localhost:8000/ws/radar)	Real-time stream between backend ↔ frontend
Environment	Docker Compose with GPU	Containerized backend + frontend, GPU acceleration for inference
Monitoring	TensorBoard	Track FPS, latency, model confidence, etc.