Skip to content

README.md

Latest commit

 

History

History
102 lines (76 loc) · 3.99 KB

README.md

File metadata and controls

102 lines (76 loc) · 3.99 KB
Raw

NET: Neuromorphic Event Triage

Robust Visual Place Recognition in High-Entropy Environments
Optimizing Ultra-Low-Energy Localization via Spatio-Temporal Noise Filtering

License: Apache 2.0 Python 3.9+

Abstract

Event cameras (Dynamic Vision Sensors) offer microsecond-level temporal resolution and ultra-low power consumption, making them ideal sensors for long-endurance autonomous navigation. However, real-world DVS deployments suffer from Event-Storm noise which are bursts of spurious events caused by lighting changes, sensor noise, and high-frequency environmental flickering that overwhelm downstream Spiking Neural Networks (SNNs) and degrade Visual Place Recognition (VPR) accuracy (Hines et al., 2025).

NET investigates a bio-inspired Spatio-Temporal Poisson Filter with an adaptive feedback controller that mimics biological retinal filtering. By dynamically triaging event streams in real-time, NET protects downstream localization models from data saturation during noise bursts, ensuring robust performance in high-entropy environments.

Key Results

NET achieves state-of-the-art noise rejection while preserving critical navigational cues.

Metric Result Description
SNR Improvement +13.82 dB Significant enhancement in Signal-to-Noise Ratio vs. raw input
Signal Preservation 99% Nearly lossless retention of structural features
Latency 81 ms Real-time processing on mobile hardware (via Numba acceleration)
Data Reduction ~70% Massive reduction in downstream bandwidth during Event-Storms

Final Comparison

Architecture

graph LR
    A["DVS Sensor<br/>(Event Stream)"] --> B["Event-Storm<br/>Detection"]
    B --> C["Spatio-Temporal<br/>Poisson Filter"]
    C --> D["Adaptive<br/>Controller"]
    D -->|"adjust k, Δt"| C
    C --> E["Filtered<br/>Event Stream"]
    E --> F["SNN / VPR<br/>Pipeline"]

    style A fill:#1a1a2e,stroke:#e94560,color:#fff
    style B fill:#1a1a2e,stroke:#e94560,color:#fff
    style C fill:#0f3460,stroke:#e94560,color:#fff
    style D fill:#0f3460,stroke:#e94560,color:#fff
    style E fill:#16213e,stroke:#0f3460,color:#fff
    style F fill:#16213e,stroke:#0f3460,color:#fff
Loading

The "Synapse" Controller

The AdaptiveEventController dynamically adjusts its neighbor threshold ($K$) using a Hysteresis Loop (Fast-Tighten, Slow-Relax). This mimics synaptic depression, preventing "filter chatter" and ensuring stability during rapid environmental changes.

Adaptive Trace

Figure: The controller (Pink) steps up $K$ when Event Density (Blue) exceeds the safety threshold, minimizing noise leak during storms.

Installation

git clone https://github.com/Cronenberg64/NET-VPR.git
cd NET-VPR
pip install -r requirements.txt

Requirements

  • Python ≥ 3.9
  • NumPy, h5py, Matplotlib, SciPy
  • Numba — JIT compilation for real-time performance

Quick Start

# Generate a synthetic noisy DVS dataset and visualize the Event-Storm
python -m src.generator

# Run the full benchmark pipeline
python -m src.main

Project Structure

NET-VPR/
├── src/
│   ├── main.py            # Benchmark Entry Point
│   ├── filter.py          # Numba-accelerated Poisson filter
│   ├── controller.py      # Adaptive feedback controller 
│   └── generator.py       # DVS event stream simulator
├── results/               # Publication-quality figures
├── data/
├── LICENSE
└── README.md

License

This project is licensed under the Apache License 2.0 — see the LICENSE file for details.