How radar-lab/mmfall works
Project type and purpose
• This is neither a web nor a traditional full-stack app. It is a sensor data-processing + machine-learning system built around ROS (Robot Operating System) to detect human falls using a 4D mmWave radar and an RNN-autoencoder anomaly detector (the HVRAE).
• The README confirms: data collection via mmWave radar (and optionally a webcam), preprocessing, semi-supervised ML model training on normal activities, and real-time inference to flag falls.
File structure
• Top-level Python scripts:
• combine.py – merging or aggregating datasets
• data_pre.py – raw → cleaned/preprocessed data
• data_analyzer.py – statistical analysis or feature exploration
• data_visualizer.py – plotting / RViz configs
• data/ – raw and processed datasets (npy, csv) plus recorded ROS bag files
• saved_model/ – trained model files (HVRAE_SL_mdl.h5, HVRAE_mdl.h5, RAE_mdl.h5)
• src/mmfall.ipynb – notebook to reproduce experiments and ROC curves
• src/ros/ – ROS workspace with three packages:
• node1_radarinterface – C++ ROS node to interface mmWave radar and publish RadarScan messages
• serial – generic serial library, used by radarinterface
• usb_webcam – optional camera capture node
Insights from the README
• Confirms a two-stage system: (i) real-time sensor acquisition via ROS, (ii) anomaly detection via Hybrid Variational RNN AutoEncoder.
• The provided misc/mmfall.png hints at the high-level data flow.
Final recommendations
• Major directories/packages as distinct components
• Note on the bifurcation: “Offline training” vs. “Online detection”