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
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





![Novel versatile hydrogen drone developed by TUDelft - MAVLab [EN]
http://www.nederdrone.nl/ Novel versatile hydrogen drone developed by TUDelft - MAVLab [EN]](https://i.ytimg.com/vi/lzvsctwbL2c/mqdefault.jpg)
![Stability-based Scale Estimation for Monocular SLAM
* Paper available at: http://ieeexplore.ieee.org/document/8246525/
* Additional demo is available at: https://youtu.be/Dp6XhvQnWp0
* Note: ORB-SLAM clip is originally from https://youtu.be/_9VcvGybsDA
[Abstract]
We propose a novel method to deal with the scale ambiguity in monocular SLAM based on control stability. We analytically show that (1) using unscaled state feedback from monocular SLAM for control can lead to system instability, and (2) there is a unique linear relationship between the absolute scale of the SLAM system and the control gain at which such instability arises. Using this property, our system estimates the scale by adapting the gain and detecting the self-induced oscillations. Unlike conventional monocular approaches, no additional metric sensors are used for scale estimation. We demonstrate the ability of our method to estimate the scale for performing autonomous indoor navigation with a low-cost quadrotor MAV. Stability-based Scale Estimation for Monocular SLAM](https://i.ytimg.com/vi/m0IvL6PD4gM/mqdefault.jpg)



