5.6. Radar Data Pre-Processing Pipeline — Complete Reference Guide

Complete guide to the Radar Data Pre-Processing Pipeline in Tensorlab

Important Note : This pipeline ONLY works on Processed Signal Chain Output and is mainly having inbuilt presets needed to preprocess Radar dataset through filterings, points/radar frames quality checks, windowing etc. and make it ready for Training


5.6.1. Table of Contents

  1. Identify Your Use Case Scenario

  2. Converting mmWave Data Recorder Captures

  3. Feature Extraction Presets

  4. Inline Override: Customizing Presets in YAML

  5. Use Case : Pre-Cleaned / Ready To Train Datasets

  6. All Configurable Parameters Reference


5.6.2. Identify Your Use Case Scenario

Your Data Situation

Recommended Path

Already have per-frame features (CSV columns = model input features, no bin*, point_, track_ columns)

Use Generic_Radar_SimpleWindowOnly preset - add inline args - frame_size, stride_size for windowing customisation

On-device Signal Chain Output from mmWave Data Recorder (columns MUST match expected names: bin*, point_x*, track_velocity_*, etc.)

Use Generic_248Input_Posture_and_Fall_Detection or Generic_392Input_PointCloudWithTrack_SimpleWindow or Generic_130Input_Radar_SurfaceClassification Presets and override params


5.6.3. Converting mmWave Data Recorder Captures

If your data comes directly from the mmWave Data Recorder, use the annotate_radar_dataset.py script to convert it into a Tensorlab-ready dataset ZIP.

Input Structure Needed (mmWave Data Recorder output):

Important Modification needed in binData Folder from User’s end : To develop ‘class label’ named folders (eg. falling, walking etc.) and classify the session folders in binData accordingly.

binData/
├── falling/
│   ├── session_001/
│   │   ├── replay_*.csv
│   │   └── replay_*.json
│   └── ...
├── sitting/
├── standing/
├── walking/
└── laying/

Command:

python examples/radar_posture_and_fall_detection/annotate_radar_dataset.py \
  --input /path/to/binData \
  --output /path/to/my_radar_dataset.zip

Output Structure (ready for Tensorlab):

my_radar_dataset.zip
└── classes/
    ├── falling/
    │   ├── session_001.csv
    │   └── ...
    ├── sitting/
    ├── standing/
    ├── walking/
    └── laying/

Then update config.yaml:

dataset:
  input_data_path: /path/to/my_radar_dataset.zip

Refer to the Radar Posture & Fall Detection README for detailed instructions specific to the application.


5.6.4. Feature Extraction Presets

5.6.4.1. Preset 1: Generic_248Input_Posture_and_Fall_Detection

Use Case: Optimized for human posture and fall detection (matches RF_DualStream_PointCloud_19K).

Generic_248Input_Posture_and_Fall_Detection = dict(
    data_processing_feature_extraction=dict(
        data_proc_transforms=['Radar Point Cloud Data', 'Radar Track Data', 'Simple Window'],
        frame_size=8,
        stride_size=0.125,
        radar_point_cloud_data_selector=['y', 'z','snr'],
        radar_point_cloud_coordinate_system='relative',
        radar_point_cloud_relative_coordinate_filter=['y'],
        radar_point_cloud_spatial_filter_bounds=[None, None, -10.0, 10.0, -10.0, 10.0],
        radar_track_data_selector=['z', 'velocity', 'acceleration'],
        radar_track_velocity_bounds=[-10.0, 10.0],
        radar_track_acceleration_bounds=[-10.0, 10.0],
        radar_track_spatial_filter_bounds=[None, None, None, None, -10.0, 10.0],
        radar_max_points_per_frame=8,
        radar_min_points_per_window=12,
        radar_point_cloud_structure_strategy='highest and lowest',
        radar_point_cloud_sort_by='z',
        radar_frames_quality_index='per_window',
        radar_feature_ordering='frame_major',
    ),
    common=dict(task_type=TASK_TYPE_RADAR_POINT_CLOUD_CLASSIFICATION),
)

Per-Frame Feature Layout (31 features):

Index

Feature

Source

0

track_height_z

Track

1

track_velocity_x

Track

2

track_velocity_y

Track

3

track_velocity_z

Track

4

track_acc_x

Track

5

track_acc_y

Track

6

track_acc_z

Track

7-30

8 points x [y, z, snr]

Point Cloud (24 features)

Total per frame: 7 (track) + 24 (8 points x 3) = 31 features
Window (8 frames): 31 x 8 = 248 features → Matches model input!


5.6.4.2. Preset 2: Generic_130Input_Radar_SurfaceClassification

Use Case: Surface Classification using FFT range bins only.

Generic_130Input_Radar_SurfaceClassification = dict(
    data_processing_feature_extraction=dict(
        data_proc_transforms=['Radar Range Data', 'Simple Window'],
        radar_range_bins_bounds=[6, 31],
        frame_size=5,
        stride_size=0.2,
        radar_frames_quality_index=None,
        variables=1,
    ),
    common=dict(task_type=TASK_TYPE_RADAR_SURFACE_CLASSIFICATION),
)

Result: 26 bins (6-31 inclusive) x 5 frames = 130 features


5.6.4.3. Preset 3: Generic_Radar_SimpleWindowOnly

Use Case: Pre-cleaned/refined datasets where CSV already has correct per-frame features. Only windowing is needed.

Generic_Radar_SimpleWindowOnly = dict(
    data_processing_feature_extraction=dict(
        data_proc_transforms=['Simple Window'],
        frame_size=8,
        stride_size=0.125,
    ),
    common=dict(task_type=TASK_TYPE_RADAR_POINT_CLOUD_CLASSIFICATION),
)

Notes:

  • No range bin, point cloud, track, or scaling stages run

  • Input CSV must already have the correct per-frame feature columns


5.6.4.4. Preset 4: Generic_392Input_PointCloudWithTrack_SimpleWindow

Use Case: Complete CSV processing with full feature extraction and quality filtering.
Resulting Input Tensor: (N, 8 x 49) = (N, 392) — 5 PC features x 8 points + 9 track features = 49/frame.

Generic_392Input_PointCloudWithTrack_SimpleWindow=dict(
    data_processing_feature_extraction=dict(
        data_proc_transforms=['Radar Point Cloud Data', 'Radar Track Data', 'Simple Window'],
        frame_size=8,
        stride_size=0.125,
        radar_point_cloud_data_selector=['x','y', 'z','doppler','snr'],
        radar_point_cloud_coordinate_system='relative',
        radar_point_cloud_relative_coordinate_filter=['x','y','z'],
        radar_point_cloud_spatial_filter_bounds=[None, None, None, None, None, None],
        radar_track_data_selector=['x', 'y', 'z', 'velocity', 'acceleration'],
        radar_track_velocity_bounds=[None, None],
        radar_track_acceleration_bounds=[None, None],
        radar_track_spatial_filter_bounds=[None, None, None, None, None, None],
        radar_max_points_per_frame=8,
        radar_min_points_per_window=12,
        radar_point_cloud_structure_strategy='highest only',
        radar_point_cloud_sort_by='z',
        radar_frames_quality_index='per_window',
        radar_feature_ordering='frame_major',
    ),
    common=dict(task_type=TASK_TYPE_RADAR_POINT_CLOUD_CLASSIFICATION), )

5.6.5. Inline Override: Customizing Presets in YAML

Power User Feature: Use a preset as base, then override specific parameters inline.

5.6.5.1. How It Works

data_processing_feature_extraction:
  feature_extraction_name: "Generic_248Input_Posture_and_Fall_Detection"
  
  # Override ANY parameter from the preset:
  frame_size: 16                    # Change window size
  stride_size: 0.25                 # Change stride
  radar_min_points_per_window: 20   # Stricter quality
  radar_max_points_per_frame: 12    # More points per frame
  radar_point_cloud_data_selector: ['x', 'y', 'z', 'snr']  # Add x coordinate

5.6.5.2. Override Example: Custom Fall Detection Config

data_processing_feature_extraction:
  feature_extraction_name: "Generic_248Input_Posture_and_Fall_Detection"
  
  # Custom: 16 frames, no overlap, stricter quality, all point features
  frame_size: 16
  stride_size: 1.0
  radar_min_points_per_window: 20
  radar_point_cloud_data_selector: ['x', 'y', 'z', 'doppler', 'snr']
  radar_point_cloud_sort_by: 'y'  
  
  # Keep rest from preset (relative coords, track features, etc.)

5.6.6. Use Case : Pre-Cleaned / Ready To Train Datasets

Scenario: You already have a dataset where each row = one frame with the exact features your model expects. No bin*, point_, track_ columns exist. You just need windowing (sliding window with stride).

5.6.6.1. When to Use

  • Features already extracted externally (MATLAB, Python, custom scripts) with no pre-processing needed

  • CSV columns = final per-frame features (e.g., 31 columns per frame)

  • No radar-specific column names present

  • You only need: frame_size, stride, no quality filtering support

5.6.6.2. Use the Preset

data_processing_feature_extraction:
  feature_extraction_name: "Generic_Radar_SimpleWindowOnly"
  # That's it! Only "Simple Window" stage runs.

5.6.6.3. Inline Config (Full Control)

data_processing_feature_extraction:
  data_proc_transforms:
    - "Simple Window"          # ONLY this stage runs
  frame_size: 8                # Frames per window (MUST match model)
  stride_size: 0.5             # Stride as fraction of frame_size (0-1)
  # Feature ordering in flattened window
  radar_feature_ordering: 'frame_major'  # or 'feature_major'

5.6.6.4. Important Notes

Parameter

Requirement

Column names

Don’t matter — pipeline treats all columns as features

Quality filtering

Only supportive when column names present as per Tensorlab standards


5.6.7. All Configurable Parameters Reference

Note: Using a preset via feature_extraction_name is highly recommended. Any parameter whose value matches the preset’s default does not need to be specified in your YAML. Only override parameters that differ from the preset defaults.

5.6.7.1. Data Processing Feature Extraction Parameters

Parameter

Type

Default

Description

feature_extraction_name

str

None

Preset Name (Eg. Generic_248Input_Posture_and_Fall_Detection for Posture and Fall Detection)

data_proc_transforms

list

[]

Main pipeline gate — tokens: “Radar Range Data”, “Radar Point Cloud Data”, “Radar Track Data”, “Scaling”, “Simple Window”

frame_size

int

1

Frames per window (critical for model input shape)

stride_size

float

0.125

Stride as fraction of frame_size (0-1)


5.6.7.2. 1. Range Data Parameters

Active only when “Radar Range Data” is in data_proc_transforms (e.g., surface classification).

Parameter

Type

Default

Description

When to Use

radar_range_bins_bounds

list[2]

[min_bin, max_bin]

Range bin selection to crop near/far bins.

Required if “Radar Range Data” in data_proc_transforms. Set to crop near/far bins. Optional: omit to keep all bins.


5.6.7.3. 2. Point Cloud Data Parameters

Active only when “Radar Point Cloud Data” is in data_proc_transforms.

Parameter

Type

Default

Description

When to Use

radar_point_cloud_data_selector

list[str]

[x, y, z, doppler, snr]

Point features to keep. Choose a subset to reduce dimensionality (e.g., y,z,snr for fall detection).

Required if “Radar Point Cloud Data” active.

radar_point_cloud_coordinate_system

str

absolute or relative

Coordinate frame. Use relative to make point cloud position-invariant.

Required if “Radar Point Cloud Data” in data_proc_transforms.

radar_point_cloud_relative_coordinate_filter

list[str]

[x, y, z]

Axes to make relative (e.g., y for distance-only).

Required when coordinate_system: relative.

radar_point_cloud_spatial_filter_bounds

list[6]

[min_x, max_x, min_y, max_y, min_z, max_z]

Spatial gate to drop points outside the sensing zone.

Optional. Use None to disable.

radar_point_cloud_non_spatial_filter_bounds

list[4]

[min_dop, max_dop, min_snr, max_snr]

Doppler/SNR gate to drop noisy or implausible points.

Optional. Use None to disable.

radar_max_points_per_frame

int

Max points per frame after selection

Cap on points kept per frame.

Required if “Radar Point Cloud Data” active.

radar_point_cloud_structure_strategy

str

Point selection strategy

Options: highest only, lowest only, highest and lowest.

Required if “Radar Point Cloud Data” active.

radar_point_cloud_sort_by

str

Feature to sort points by: x, y, z, doppler, snr

Determines sort order for the point selection strategy.

Required if “Radar Point Cloud Data” active.


5.6.7.4. 3. Track Data Parameters

Active only when “Radar Track Data” is in data_proc_transforms.

Parameter

Type

Default

Description

When to Use

radar_track_data_selector

list[str]

[x, y, z, velocity, acceleration]

Track features to keep. Select subset to control features/frame (e.g., z, velocity, acceleration).

Required if “Radar Track Data” active.

radar_track_velocity_bounds

list[2]

[min_vel, max_vel]

Filter out physically implausible track velocities (ghost tracks). Eg: [-10, 10].

Optional.

radar_track_acceleration_bounds

list[2]

[min_acc, max_acc]

Filter out implausible accelerations. Eg: [-10, 10].

Optional.

radar_track_spatial_filter_bounds

list[6]

[min_x, max_x, min_y, max_y, min_z, max_z]

Gate track position to the sensing zone.

Optional. Use None to disable.

radar_data_scaling_track_velocity

float

Scale factor for track_velocity_* columns

Independent of “Radar Track Data”. Use when velocity magnitude is too small vs position features (e.g., 30).

Optional. You NEED to specify “Scaling” pre processing token in data_proc_transforms


5.6.7.5. 4. Simple Window Parameters

Active only when “Simple Window” is in data_proc_transforms.

Parameter

Type

Default

Description

When to Use

frame_size

int

1

Frames per window. Critical for model input shape.

Always when windowing.

stride_size

float

0.125

Stride as fraction of frame_size (0-1).

Always when windowing.


5.6.7.6. 5. Quality Indexing Parameters

Controls frame/window-level quality filtering.

Parameter

Type

Default

Description

When to Use

radar_frames_quality_index

str

per_radar_frame, per_window, or unset

Quality filtering mode. None = disabled. per_radar_frame works standalone. per_window needs “Simple Window”.

Optional

radar_max_points_per_frame

int

Max points per frame after selection

Cap on points kept per frame (shared with point cloud).

Required if “Simple Window”, “Point Cloud Data” active and Quality filtering needed.

radar_min_points_per_frame

int

Min real points required per frame (per_radar_frame mode)

Drop individual frames with < N real points.

Required only if radar_frames_quality_index: per_radar_frame. Set >0 to drop sparse frames. Make sure “Point Cloud Data” is active in data_proc_transforms

radar_min_points_per_window

int

Min total real points across window (per_window mode)

Drop entire windows with < N total points across all frames.

Required only if radar_frames_quality_index: per_window. Set >0 to drop low-quality windows. Make sure “Point Cloud Data” is active in data_proc_transforms

Comparison:

Mode

Parameters

Drops

Requires “Simple Window”

unset / empty

N/A

Nothing

No

per_radar_frame

radar_min_points_per_frame

Individual frames with < N real points

No (works standalone)

per_window

radar_min_points_per_window

Entire windows with < N total points across all frames

Yes


5.6.7.7. 6. Feature Ordering Parameter

Parameter

Type

Default

Description

When to Use

radar_feature_ordering

str

frame_major or feature_major

Flattening order for the window tensor. Must match the model’s expectation.

Optional. Set when the model expects a specific feature layout.


Generated for Tensorlab EdgeAI — Making mmWave Radar ML Accessible