Uploaded April 2021 | Updated November 2026, 3 hours ago
Video demonstrating the performance of the pose estimation network and network-in-loop autonomous flights. A part of this work (shown during the first 1 minute and 35 seconds of this video) has been accepted by ICRA 2021.
Abstract: In the field of visual ego-motion estimation for Micro Air Vehicles (MAVs), fast maneuvers stay challenging mainly because of the big visual disparity and motion blur. In the pursuit of higher robustness, we study convolutional neural networks (CNNs) that predict the relative pose between subsequent images from a fast-moving monocular camera facing a planar scene. Aided by the Inertial Measurement Unit (IMU), we mainly focus on the translational motion. The networks we study have similar small model sizes (around 1.35MB) and high inference speeds (around 100Hz on a mobile GPU). Images for training and testing have realistic motion blur. Departing from a network framework that iteratively warps the first image to match the second with cascaded network blocks, we study different network architectures and training strategies. Simulated datasets and MAV flight datasets are used for evaluation. The proposed setup shows better accuracy over existing networks and traditional feature-point-based methods during fast maneuvers. Moreover, self-supervised learning outperforms supervised learning. In the Appendix, we further demonstrate the network's performance by testing on a public high-speed flight dataset and running it in real-time onboard for the feedback control of an autonomous MAV that reaches speeds up to 4m/s.
Article (with Appendix): arxiv.org/abs/2101.01841
Open-Sourced Code: github.com/tudelft/PoseNet_Planar
Authors: Yingfu Xu, Prof. Dr. Guido C. H. E. de Croon
Technical Support: Ir. Nilay Y. Sheth
The narrator of this video is generated using Google Cloud Text-to-Speech API.
Video demonstrating the performance of the pose estimation network and network-in-loop autonomous flights. A part of this work (shown during the first 1 minute and 35 seconds of this video) has been accepted by ICRA 2021.
Abstract: In the field of visual ego-motion estimation for Micro Air Vehicles (MAVs), fast maneuvers stay challenging mainly because of the big visual disparity and motion blur. In the pursuit of higher robustness, we study convolutional neural networks (CNNs) that predict the relative pose between subsequent images from a fast-moving monocular camera facing a planar scene. Aided by the Inertial Measurement Unit (IMU), we mainly focus on the translational motion. The networks we study have similar small model sizes (around 1.35MB) and high inference speeds (around 100Hz on a mobile GPU). Images for training and testing have realistic motion blur. Departing from a network framework that iteratively warps the first image to match the second with cascaded network blocks, we study different network architectures and training strategies. Simulated datasets and MAV flight datasets are used for evaluation. The proposed setup shows better accuracy over existing networks and traditional feature-point-based methods during fast maneuvers. Moreover, self-supervised learning outperforms supervised learning. In the Appendix, we further demonstrate the network's performance by testing on a public high-speed flight dataset and running it in real-time onboard for the feedback control of an autonomous MAV that reaches speeds up to 4m/s.
Article (with Appendix): arxiv.org/abs/2101.01841
Open-Sourced Code: github.com/tudelft/PoseNet_Planar
Authors: Yingfu Xu, Prof. Dr. Guido C. H. E. de Croon
Technical Support: Ir. Nilay Y. Sheth
The narrator of this video is generated using Google Cloud Text-to-Speech API.





![[Additional Demo] Stability-based Scale Estimation for Monocular SLAM
* Paper is available at: http://ieeexplore.ieee.org/document/8246525/
* The main demo is at: https://youtu.be/m0IvL6PD4gM
In this video, we show two autonomous flights:
1. We run our stability-based scale estimation while hovering at a small distance to the scene (2 m). Afterwards, the MAV flies backward for 10 m, hovers for 5 seconds, then comes back to the initial position and hovers. We show that the MAV hovers stably even at a large scene distance (see 2:50) and comes back to the start position accurately (compare the drones location at 0:55 and 1:40). We also display the control input and real-time SLAM output.
2. We do the similar flight, but this time we estimate the scale at a large distance to the scene (over 12 m), then fly forward and backward for 10 m. We show that the MAV hovers stably even at a small scene distance (see 5:25) and comes back to the start position (compare 3:40 and 4:19, especially the onboard view).
This demonstrates that the proposed system can still be used in a larger scale scenario with very large scene distances and variations. [Additional Demo] Stability-based Scale Estimation for Monocular SLAM](https://i.ytimg.com/vi/Dp6XhvQnWp0/mqdefault.jpg)




