| name | aquatic-neuromorphic-optical-flow |
| description | Self-supervised spiking neural network framework for estimating per-pixel optical flow from event camera streams in underwater environments. Bridges neuromorphic sensing and aquatic intelligence for lightweight, real-time, low-cost perception on resource-constrained edge platforms. Activation: underwater vision, event camera, neuromorphic optical flow, aquatic perception, spiking neural network motion estimation, DVS underwater. |
| category | neuroscience |
Aquatic Neuromorphic Optical Flow
Self-supervised SNN framework for per-pixel optical flow estimation from asynchronous event streams in underwater environments, bypassing the underwater data scarcity bottleneck.
Metadata
- Source: arXiv:2605.07653v1
- Authors: Pei Zhang, Yunkai Liang, Kaiqiang Wang
- Published: 2026-05-08
- Categories: cs.CV, eess.IV
- Status: Under review
Core Methodology
Problem
Underwater imaging faces severe constraints: light attenuation, scattering, turbidity, and strict resource limits on autonomous underwater vehicles (AUVs). Conventional frame-based cameras produce redundant data in aquatic environments where most of the scene is static between frames.
Key Innovation: Neuromorphic Vision for Aquatic Perception
- Event cameras (Dynamic Vision Sensors) report only pixel-level brightness changes (asynchronous events)
- Data bandwidth reduction: 10-100x compared to conventional RGB video
- High temporal resolution: microsecond-level event timestamps capture fast underwater motion
- High dynamic range: handles extreme lighting transitions underwater
Self-Supervised SNN Framework
- Input: Asynchronous event streams from DVS cameras in underwater scenarios
- Spiking Neural Network: Encodes event spatiotemporal patterns into spike trains
- Motion field estimation: SNN learns to predict per-pixel optical flow without ground-truth labels
- Self-supervision: Uses event warping consistency — predicted flow should align events to form a coherent image
- Output: Dense optical flow field for downstream tasks (navigation, obstacle avoidance, object tracking)
Self-Supervision via Event Warping
- Predict flow field that "warps" events backward in time
- Minimize reconstruction error of warped events (events should align if flow is correct)
- No need for labeled optical flow data, bypassing the underwater annotation bottleneck
Implementation Guide
Prerequisites
- Event camera data (DVS) from underwater scenarios
- SNN training framework (SpikingJelly, Lava, or custom)
- GPU for training
Architecture
Events (x, y, t, polarity)
→ Voxel grid representation (temporal binning)
→ SNN encoder (spiking conv layers)
→ Flow decoder
→ Optical flow field (u, v per pixel)
→ Event warping + reconstruction loss
→ Self-supervised training loop
Key Steps
- Convert asynchronous events to voxel grid representation with temporal bins
- Build SNN with leaky integrate-and-fire (LIF) neurons for temporal encoding
- Train with event warping self-supervision loss
- Evaluate optical flow quality via downstream task performance
Applications
- AUV navigation: Real-time motion perception for autonomous underwater vehicles
- Underwater obstacle avoidance: Low-latency collision detection
- Marine biology monitoring: Tracking aquatic organisms with minimal power
- Subsea inspection: Pipeline, cable, and structure monitoring
- Resource-constrained edge platforms: Deploy on battery-powered underwater sensors
Related Skills
- neuromorphic-spinnaker-asl
- snn-near-sensor-noise-filter-dvs
- direct-to-event-snn-transfer
- event-driven-neuromorphic-transceiver
- neuromorphic-spacecraft-pose-event-camera
