NanoFlowNet: Real-Time Dense Optical Flow on a Nano Quadcopter @microuav
NanoFlowNet: Real-Time Dense Optical Flow on a Nano Quadcopter  @microuav
Uploaded May 2023 | Updated November 2026, 4 days ago
We have for the first time deployed a fully convolutional net, called “NanoFlowNet”, on CrazyFlie’s AI deck. In order to succeed, we redesigned a small fully convolutional network from the image segmentation literature for optical flow estimation. Simultaneously, we optimized for reduced operations, while retaining as much flow estimation performance as possible. We found that the performance is better if the network is explicitly guided to retain motion edges in the field of view.

The network runs onboard of the CrazyFlie at 5.57-9.34 frames per second. The drone used the flow to avoid obstacles in our Cyberzoo. Inspired by the GapFlyt work of @Yiannis Aloimonos’ group at @Maryland university, we had the drone actively move up and down to detect obstacles better.

Paper: arxiv.org/pdf/2209.06918.pdf
NanoFlowNet: Real-Time Dense Optical Flow on a Nano QuadcopterIMAV 2014 Presentations - Fabrizio SchianoIMAV2015 MAVlab TUDelftTU Delft Quadshot - Proof of FlightEnd-to-end Neural Network Based Optimal Quadcopter ControlIMAV 2022 - Greenhouse ChallengeNovel versatile hydrogen drone developed by TUDelft - MAVLab [EN]Stability-based Scale Estimation for Monocular SLAMLisaS on quad and HeliIMAV 2014 Presentations - Sophie ArmaniniPaparazzi On-screen DisplayAR.Drone 2 Autonomous flight with Paparazzi
MAVLab TU Delft |

NanoFlowNet: Real-Time Dense Optical Flow on a Nano Quadcopter

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER