7.10.40. Radar Surface Classification
Classify the surface or material in front of a TI mmWave radar (for example grass vs not grass) directly from 1-D FFT range-profile data on the xWRL6432. No point cloud or tracker is needed: the shape of the energy return across range bins is the feature. New in 1.6.0.
Example folder: examples/radar_surface_classification
(GitHub).
The folder has its own detailed README and config.yaml.
7.10.40.1. Overview
Task:
radar_surface_classification(target_module: radar)Device: xWRL6432 (Arm Cortex-M4F, no NPU)
Dataset:
radar_grass_surface_dataset.zip(grass vs not grass), downloaded by the configFeature extraction: preset
Generic_130Input_Radar_SurfaceClassification: range bins 6-31 (26 bins) x 5 frames = 130 inputs,stride_size: 0.2Model:
RF_MLP_3L_11K(10,982 parameters at 130 inputs, 2 classes).RF_MLP_4L_19Kalso works and adapts its first layer to the input size (11,510 parameters at 130 inputs).
7.10.40.2. Running the Example
./run_tinyml_modelzoo.sh examples/radar_surface_classification/config.yaml
7.10.40.3. Configuration
common:
task_type: radar_surface_classification
target_module: radar
target_device: xWRL6432
dataset:
dataset_name: radar_surface_classification
input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/radar_grass_surface_dataset.zip
split_type: amongst_files
split_factor: [0.7, 0.2, 0.1]
data_processing_feature_extraction:
feature_extraction_name: Generic_130Input_Radar_SurfaceClassification
frame_size: 5
stride_size: 0.2
radar_range_bins_bounds: [null, null]
training:
model_name: RF_MLP_3L_11K
training_epochs: 100
learning_rate: 0.001
optimizer: SGD
output_int: false
num_gpus: 0
testing: {}
compilation: {}
radar_range_bins_bounds: [null, null]keeps all range bins; the preset default is[6, 31]. Changing the bins orframe_sizechanges the number of model inputs.Input CSV columns are
bin0…binN, one row per frame. See Radar Data Pre-Processing Pipeline — Complete Reference Guide.split_type: amongst_fileskeeps whole recordings in one split.On the device, the same range-bin selection and windowing must be applied to the live FFT output before calling
tvmgen_default_run().