This research proposes AdaThresh-EVS, a novel adaptive thresholding framework for event-based vision sensors to address limitations of fixed-threshold spike detection. The method integrates multi-scale temporal context analysis, histogram-based percentile estimation for adaptive threshold computation, and spatial coherence refinement to dynamically adjust detection thresholds based on local spatiotemporal statistics. The framework aims to improve signal preservation while rejecting background activity noise across varying illumination conditions, with validation planned on E-MLB, LED, and DVS-NOISE20 benchmarks against EDnCNN, MLPF, and SA-filter baselines. Designed for computational efficiency, the system targets real-time edge deployment on neuromorphic hardware.
Key findings
The proposed AdaThresh-EVS framework dynamically adjusts detection thresholds based on local spatiotemporal statistics, addressing the static inadequacy of fixed-threshold approaches.
Multi-scale temporal context analysis enables robust characterization of background activity noise under varying illumination and motion dynamics.
Histogram-based percentile estimation allows pixel-specific adaptive threshold computation that balances sensitivity and noise rejection.
Spatial coherence constraints in the refinement stage enforce consistency across local neighborhoods to filter isolated noise events.
The architecture achieves computational efficiency suitable for real-time edge deployment on neuromorphic hardware while maintaining detection latency.
Limitations & open questions
Validation remains pending on proposed benchmarks; actual performance gains are theoretical until empirical testing is completed.
Adaptive thresholding may introduce computational overhead compared to fixed-threshold methods, requiring careful optimization for resource-constrained edge devices.
The method assumes spatial coherence of signal events, which may not hold for extremely sparse or isolated spike patterns.