7.9.39. Radar Pose and Fall Detection

Classify the posture of a person (standing, sitting, lying, falling, walking) from mmWave radar point cloud data on the IWRL6432. This is the radar point cloud classification example. New in 1.5.0.

Example folder: examples/radar_pose_and_fall_detection (GitHub)

7.9.39.1. Overview

  • Task: radar point cloud classification (radar_classification, target_module: radar), 5 classes

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

  • Device: IWRL6432 (config.yaml, the only config in the folder). It is the only target device of the radar module; the compiler toolchain is ti-cgt-armllvm for the Cortex-M4F.

  • Model: Pose_and_Fall_model, a four-layer fully connected network (BatchNorm and Linear layers 64, 32, 16 and 5 outputs, implemented as LINEAR_4L_PC in tinyml_modelzoo/models/radar.py)

7.9.39.2. The Dataset

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

Index

Class

0

STANDING

1

SITTING

2

LYING

3

FALLING

4

WALKING

Each CSV row is a radar point cloud frame with the tracker fields (posx, posy, posz, velx, vely, velz, accx, accy, accz).

7.9.39.3. Preparing Your Own Data (Optional)

pose_and_fall_dataset.py is an optional preprocessing script for users who do not use the public dataset. Put your CSV files in class folders under dataset/classes/ in the working directory, set DATASET_DIR in the script and run it. It:

  • adds missing tracker columns with a default of 0 and labels each row with its class,

  • drops rows with fewer than MIN_POINTS = 5 points or an invalid track ID,

  • filters points by height (-4 to 3 m) and distance (-4 to 5 m),

  • keeps the highest and lowest points of each point cloud frame as the model input,

  • writes the cleaned data to classes_cleaned/ and creates instances_train_list.txt and instances_test_list.txt with an 80/20 split.

7.9.39.4. Running the Example

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

./run_tinyml_modelzoo.sh examples/radar_pose_and_fall_detection/config.yaml

7.9.39.5. Configuration

common:
  task_type: radar_classification
  target_module: radar
  target_device: IWRL6432
dataset:
  dataset_name: radar_human_pose_detection
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/radar_human_pose_detection.zip
  split_type: within_files
  split_factor: [0.8, 0.1, 0.1]
data_processing_feature_extraction:
  data_proc_transforms:
    - SimpleWindow
  frame_size: 8
  stride_size: 0.5
training:
  model_name: Pose_and_Fall_model
  batch_size: 64
  training_epochs: 160
  early_stopping: true
  early_stopping_patience: 5
  num_gpus: 0
  quantization: 0
  learning_rate: 0.0001
  optimizer: SGD
testing: {}
compilation: {}

  • data_dir is detected automatically as classes for radar classification.

  • split_type: within_files splits every file into 80% train, 10% validation and 10% test windows.

  • SimpleWindow with frame_size: 8 groups 8 consecutive point cloud frames per sample.

  • quantization: 0 trains a float model (the compilation stage uses the software path for float models).

7.9.39.6. Results

The example folder has no readme with accuracy numbers. In the automated 1.5.0 config test run (test set of 1999 windows: 380 falling, 324 lying, 491 sitting, 544 standing, 260 walking) the float model reached:

Metric

Value

Best epoch (of 160, early stopping enabled)

141

Validation accuracy / F1 score (best epoch)

96.27% / 0.963

Test accuracy

96.10%

Test AUC ROC

1.000

7.9.39.7. Output Location

Results are written to ./data/projects/radar_human_pose_detection/run/<date-time>/Pose_and_Fall_model/. The trained model is in training/base and the compiled model in compilation. See Understanding the Run Output and Radar Classification for the radar task type.

7.9.39.8. See Also