Tuning hampel_threshold for Noisy Environments in Micro-ESPectre

Increase HAMPEL_THRESHOLD from its default 5.0 to 6.0–7.0 in noisy RF environments while optionally enlarging HAMPEL_WINDOW to 9, ensuring the filter removes interference spikes without suppressing genuine motion signatures.

Micro-ESPectre processes Wi-Fi Channel State Information (CSI) to detect motion through spatial turbulence analysis. The HampelFilter class removes isolated outliers from the turbulence signal that would otherwise trigger false motion reports in the francescopace/espectre repository. Tuning the hampel_threshold parameter is critical when deploying in environments with strong RF interference from multiple Wi-Fi devices or industrial equipment.

How the Hampel Filter Processes CSI Turbulence

The filter maintains a circular buffer of the last window samples. For each new turbulence sample, it computes the median of the buffer and the Median Absolute Deviation (MAD). If the sample deviates from the median by more than threshold × MAD, the filter replaces it with the median; otherwise, it passes through unchanged. This robust statistical approach isolates spikes without affecting the underlying variance that the MVS (Motion Vector Segmentation) algorithm requires.

Configuration Parameters and Default Values

Default values are defined in micro-espectre/src/config.py at lines 76-80:

# config.py

HAMPEL_WINDOW = 7
HAMPEL_THRESHOLD = 5.0

These constants flow into the MVSDetector class defined in micro-espectre/tools/csi_utils.py (lines 14-16). The constructor accepts hampel_window and hampel_threshold as arguments:

# csi_utils.py - MVSDetector initialization

def __init__(self, ..., enable_hampel=True, hampel_window=7, hampel_threshold=5.0):

The detector forwards these values to a SegmentationContext, which instantiates the HampelFilter internally (see the call at line 44 in csi_utils.py). During processing, the add_turbulence method in micro-espectre/src/segmentation.py invokes self.hampel.filter(value) to clean each incoming sample.

Tuning Strategies for Noisy Environments

Understanding the Threshold Trade-off

A lower HAMPEL_THRESHOLD makes the filter more aggressive, risking the replacement of genuine motion-related spikes and reducing detection sensitivity. A higher threshold becomes more permissive, allowing noise spikes to pass through and increasing false positives.

Recommended Configurations by Noise Level

• Mild noise, occasional spikes: Retain the defaults (HAMPEL_WINDOW=7, HAMPEL_THRESHOLD=5.0).
• Moderate interference: Increase HAMPEL_THRESHOLD to 6.0 and test HAMPEL_WINDOW=9 for smoother median calculation.
• Heavy RF noise: Raise both parameters to HAMPEL_WINDOW=9 and HAMPEL_THRESHOLD=7.0, then monitor the false-positive rate.
• Low-latency requirements: Keep HAMPEL_WINDOW minimal (≥5) and compensate by raising HAMPEL_THRESHOLD if false positives rise.

Practical Implementation and Optimization

Instantiate a Detector with Custom Hampel Settings

from micro_espectre.tools.csi_utils import MVSDetector
from micro_espectre.src.config import DEFAULT_SUBCARRIERS

detector = MVSDetector(
    window_size=100,
    threshold=2.5,
    selected_subcarriers=DEFAULT_SUBCARRIERS,
    enable_hampel=True,
    hampel_window=9,      # Smoother median for noisy environments

    hampel_threshold=7.0,  # More tolerant outlier detection

    enable_lowpass=False
)

detector.process_packet(csi_packet)

Automated Parameter Optimization

Use the optimization script at micro-espectre/tools/6_optimize_filter_params.py (lines 78-84) to search for the best configuration:

python micro-espectre/tools/6_optimize_filter_params.py --hampel

The script evaluates window sizes from 3-11 and thresholds from 2.0-6.0, outputting the configuration with the highest F-score while maintaining low false-positive rates.

Validation Metrics

After tuning, verify performance using the test_mvs_configuration helper in csi_utils.py. Ensure:

• Recall remains ≥ 95% (motion packets detected)
• False-positive rate stays ≤ 10%

If metrics deviate, adjust the threshold in steps of 0.5 until achieving the desired trade-off.

Summary

• The Hampel filter in Micro-ESPectre uses a sliding median and MAD calculation to remove outlier spikes from Wi-Fi CSI turbulence data.
• Default parameters (HAMPEL_WINDOW=7, HAMPEL_THRESHOLD=5.0) suit typical indoor conditions but require adjustment in noisy RF environments.
• Increase hampel_threshold to 6.0–7.0 for heavy interference, optionally enlarging the window to 9 for median stability.
• Validate changes using the 6_optimize_filter_params.py script or manual testing to ensure recall ≥ 95% and false positives ≤ 10%.
• Configuration originates in config.py (lines 76-80) and propagates through MVSDetector in csi_utils.py to the SegmentationContext class.

Frequently Asked Questions

What is the default hampel_threshold in Micro-ESPectre?

The default value is 5.0, defined in micro-espectre/src/config.py at line 80. This represents a multiplier of the Median Absolute Deviation (MAD) that determines the outlier detection limit.

How does increasing the Hampel window size affect detection latency?

Larger window sizes (e.g., 9 instead of 7) increase the median calculation's stability but introduce slight latency because the filter must buffer more samples. For low-latency requirements, keep HAMPEL_WINDOW at 5 or 7 and compensate by raising HAMPEL_THRESHOLD instead.

Can I disable the Hampel filter entirely?

Yes. Set enable_hampel=False when constructing the MVSDetector class in micro-espectre/tools/csi_utils.py. However, disabling the filter is not recommended for noisy environments as it will significantly increase false motion detections from RF interference spikes.

Where does the actual Hampel filtering occur in the codebase?

The filtering happens in micro-espectre/src/segmentation.py within the SegmentationContext class. The add_turbulence method calls self.hampel.filter(value) each time a new turbulence sample is processed, replacing outliers with the window median according to the configured threshold.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}