NightSight enables drones to navigate in complete darkness using event cameras
Combines a coded aperture lens and IR dot projector for real-time depth sensing in pitch black.
Get AI news that actually matters
One email a day. Zero fluff. Join 10,000+ professionals.
Researchers have unveiled NightSight, a novel passive computation approach for drone navigation in complete darkness. The system combines a monocular event camera with a coded aperture lens and an infrared dot projector. When the projected pattern passes through the coded aperture, it creates depth-dependent blur signatures that implicitly encode scene geometry — no active depth calculation needed. A convolutional neural network trained exclusively on synthetic planar wall data decodes these blur signatures into dense depth maps, and remarkably generalizes zero-shot to complex real-world scenes.
The entire pipeline runs in real time at 20 Hz on an NVIDIA Jetson Orin Nano, making it ideal for resource-constrained platforms like small aerial robots in search-and-rescue operations. Accuracy is impressive: l1 error of 7.0 cm up to 2.5 meters (2.80% relative error). The work also analyzes how different coded aperture designs affect performance. By eliminating the need for heavy LIDAR or structured light scanners, NightSight opens a path to low-power, low-weight night vision autonomy for drones in hazardous, confined environments.
- Passive technique uses a coded aperture lens to infer depth from event camera images without active scanning.
- CNN trained only on synthetic data generalizes zero-shot to cluttered real-world scenes.
- Achieves 7 cm depth error at 2.5 m range while running at 20 Hz on an NVIDIA Jetson Orin Nano.
Why It Matters
Enables small, cheap drones to autonomously navigate pitch-black disaster zones without heavy sensors or LIDAR.