Comparison of Anomaly Detection Methods on Event-Based Vision Sensor Data in a High Noise Environment

Area of Science:

• Computer Vision
• Signal Processing
• Sensor Technology

Background:

• Event-based vision sensors (EVSs) offer high temporal resolution and unique event data output.
• Anomaly detection in hyperspectral analysis can be adapted for EVS event frequency spectra.
• Strong interfering sources can significantly reduce EVS sensitivity and obscure targets.

Purpose of the Study:

• To compare five eigenanalysis anomaly detection methods for target detection in high noise EVS data.
• To evaluate detector performance in the presence of overwhelming interfering sources.
• To identify optimal detection methods for challenging EVS applications.

Main Methods:

• Applied Principal Component Background Suppression (PCBS), Mahalanobis Distance (MD), Complementary Subspace Detector (CSD), Reed-Xiaoli (RX), and Subspace Reed-Xiaoli (SSRX) detectors.
• Utilized frequency analysis, background suppression, and statistical filtering techniques.
• Evaluated detection probability against false-alarm probability in simulated high noise conditions.

Main Results:

• PCBS, MD, and CSD detectors successfully detected targets through strong interference.
• PCBS demonstrated superior performance at low false-alarm rates (e.g., >400% detection increase at 10^-5 false-alarm probability).
• MD and CSD excelled at higher false-alarm rates (approx. 7 × 10^-2), with MD offering sub-second execution.

Conclusions:

• PCBS and MD detectors are recommended for target detection in high noise EVS environments.
• The choice between PCBS and MD depends on the specific application's requirements for false-alarm rate and speed.
• Eigenanalysis anomaly detection methods show promise for robust target identification with EVS technology.