Optical flow for monocular obstacle avoidance

[leshdaemon]

Investigate whether optical flow from a single RGB camera can be used for real-time obstacle avoidance.

Autoresearch

you can run a go2 lidar + image replay and autoresearch metric is how much optical flow obstacle detecion corelates to lidar detected obstacles

theory:

• Time-to-contact (tau) from flow divergence - proven biological principle (insect navigation). Detects approaching obstacles without absolute depth.
• Sparse optical flow (Lucas-Kanade) is cheap on embedded hardware; dense flow (RAFT, FlowNet) gives better coverage at higher compute cost.
• Expanding flow patterns reliably indicate obstacles ahead for a forward-moving robot.

check

• What are the other strategies for keypoint detection that we can use for optical flow? like what does ORB-SLAM use?

Can be tested on Tello drone or Go2.

---

Synced from DIM-791 by summer