A moving object leaves information in light even when fog prevents a conventional camera from forming a clear picture. Brown University engineers recovered some of that information by pairing an event camera, which records changes in brightness, with a spiking neural network that reconstructs the hidden target.[1][2]
The laboratory system both imaged and tracked targets behind drifting fog and turbid water. Its reconstructed images reached structural-similarity scores of up to 96% against the reference targets. That is an image-resemblance metric under the study’s test conditions; it describes neither a general object-detection success rate nor a road-driving trial.[1][2]
The key change happens before image interpretation. Instead of asking AI to sharpen an ordinary photograph, the researchers use a sensor that preserves the timing of tiny changes in the arriving light.
Why motion can become an advantage
How the fog-imaging system works
| Component | Role |
|---|---|
| Dynamic vision sensor | Records pixel-level brightness changes as events |
| Spiking neural network | Uses event timing to reconstruct the target |
| Laboratory evaluation | Compares reconstructed images and tracked positions with known moving targets |
In ordinary photography, movement through mist tends to be a nuisance. A long exposure smears the moving subject; a short exposure may leave too little light. Event cameras invert the instinct. They tend to transmit changes, rather than everything that sits still. A moving boundary can therefore supply structured information even when a steady photograph is overwhelmed by scattered background light.[1][2]
In a spiking neural network, computation proceeds through discrete signalling events. That design evokes one aspect of neural activity, although it would be misleading to call the system a human eye or brain. The model was trained to recover the object’s appearance and location from those event streams, using the temporal structure that a standard image may average away.
Consider a hidden moving letter. The camera cannot simply read its printed outline as if the mist were absent. Instead, small brightness changes arrive in time as the target shifts. The algorithm must infer how those fleeting signals fit an underlying form. That is why the team's choice to reconstruct objects that move randomly is significant. A system that only handles a carefully choreographed left-to-right sweep would be less useful.[1][2]
From a controlled target to an outdoor scene
Laboratory fog and cloudy water let the researchers control the targets, distances and reference images. Outdoors, illumination, scattering and contrast vary together, while a robot must also handle stationary hazards. Those are the conditions that determine where this approach could complement existing sensors.
Designing the measurement with the model
The result shows how changing the measurement can unlock a problem that appears intractable in a conventional frame. The event stream carries timing information; the network uses that information to infer shape and motion. Larger image models alone would not reproduce that change in what the camera records.[1][2]
A useful next demonstration would retain the reconstruction and tracking advantage on unfamiliar objects outdoors. For now, the laboratory result offers a clear engineering insight: motion can supply structure that scattering hides from a still image.
Sources & further reading
- Neuromorphic Optical Tracking and Imaging of Randomly Moving Targets Through Dynamic Dense Scattering Media Advanced Science · 24 Aug 2026
- Brown University engineers build brain-inspired imaging system that can see through fog Brown University · 1 Oct 2026 · Accessed 26 Oct 2026