7.10.39. Radar Posture and Fall Detection

Classify the posture of a person (falling, laying, sitting, standing, walking) from mmWave radar point cloud and track data on the xWRL6432. Introduced in 1.5.0 as radar_pose_and_fall_detection; renamed, reworked and documented in detail in 1.6.0.

Example folder: examples/radar_posture_and_fall_detection (GitHub). The folder has its own detailed README, config.yaml and annotate_radar_dataset.py.

7.10.39.1. Overview

  • Task: radar_point_cloud_classification (target_module: radar), 5 classes

  • Application: human posture recognition and fall detection from radar point clouds, for example for elder care and in-home monitoring

  • Device: xWRL6432 (Arm Cortex-M4F, no NPU; compiled with ti-cgt-armllvm)

  • Feature extraction: preset Generic_248Input_Posture_and_Fall_Detection: 8 frames x 31 features per frame = 248 inputs (track height, velocity x/y/z, acceleration x/y/z and 8 highest/lowest points of y, z, snr)

  • Model: RF_DualStream_PointCloud_19K (about 18.9K parameters at 248 inputs): a position stream, a velocity stream and a fusion MLP. Velocity features must stay at slots 1-3 of every frame, which the preset guarantees. RF_MLP_4L_19K and RF_MLP_3L_11K accept any feature order.

7.10.39.2. The Dataset

The config downloads posture_and_fall_detection_dataset.zip from https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/. If the download fails on a corporate network, download it manually and set dataset.input_data_path to the extracted folder. Data is organised in class folders (classes/), one folder per posture:

Index

Class

0

Falling

1

Laying

2

Sitting

3

Standing

4

Walking

Column names must match the radar signal chain output (point_x*, track_velocity_* and so on). Use annotate_radar_dataset.py in the example folder to annotate your own recordings, and Radar Data Pre-Processing Pipeline — Complete Reference Guide for the column reference and every pre-processing parameter.

7.10.39.3. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/radar_posture_and_fall_detection/config.yaml

7.10.39.4. Configuration

common:
  task_type: radar_point_cloud_classification
  target_module: radar
  target_device: xWRL6432
dataset:
  dataset_name: radar_point_cloud_classification
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/posture_and_fall_detection_dataset.zip
  split_type: within_files
  split_factor: [0.7, 0.2, 0.1]
data_processing_feature_extraction:
  feature_extraction_name: "Generic_248Input_Posture_and_Fall_Detection"
  data_proc_transforms: ['Radar Point Cloud Data', 'Radar Track Data', 'Simple Window', 'Scaling']
  radar_data_scaling_track_velocity: 30
training:
  model_name: RF_DualStream_PointCloud_19K
  training_epochs: 250
  learning_rate: 0.001
  optimizer: adamw
  weight_decay: 0
  lr_scheduler: CosineAnnealingLR
  num_gpus: 0
testing: {}
compilation: {}
  • split_type: within_files splits every file into 70% train, 20% validation and 10% test windows.

  • The preset sets frame_size: 8 and stride_size: 0.125; override them only if your use case differs.

  • Quantization is off by default; set training.quantization: 1 to enable it.

7.10.39.5. Output Location

Results are written to ./data/projects/radar_point_cloud_classification/run/<date-time>_RF_DualStream_PointCloud_19K/. The trained model is in training/base and the compiled model in compilation. See Understanding the Run Output and Radar Classification.

7.10.39.6. See Also