Deep Scale Learning for Visual Motion Estimation (MA/SA)


A2e27ddf 06bd 40db b571 4f8d5da70c4c.png 500 500

Organization: Autonomous Systems Lab

Involved Host(s): Gilitschenski Igor , Schneider Thomas

Abstract: The goal of this project is to train a deep neural network to estimate scale for a purely visual motion estimation framework. This will enable robots to perform metric ego-motion estimation without use of additional information sources such as an IMU or wheel encoders.

Description: ***Have a look here https://goo.gl/xLnMWI for a nicer version.*** When motion is estimated purely from monocular visual data, it is not possible to infer a metric scale without any prior knowledge or assumptions on the environment. Therefore, typically additional sensors, such as IMUs or wheel encoders, are included into the estimation process. The goal of this project is to develop and train a deep neural network to estimate scale solely based on the information coming from the camera. Such a system will enable robots to perform metric ego-motion estimation without the use of additional information sources such as an IMU or wheel encoders or can be used as a fall-back information source in case of sensor-faults or during bad operating conditions (e.g. vibrations). ***We offer the opportunity to*** - work with a state-of-the-art visual-inertial SLAM system, - work at a world renowned robotic research lab with strong ties to major industrial partners, - work with Google Tango devices, - submit your work for publication in case of project success.

Work Packages: 1. Review of existing literature on this topic 2. Develop a network architecture to estimate scale to aid a VO system (network in-/outputs, structure, ...) 3. Create training and validation datasets using the existing visual-inertial SLAM system 4. Train/evaluate the network on the generated datasets 5. Optionally: integrate the network into the existing VIO pipeline 6. Optionally: demonstrate the method on one of our robots

Requirements: - Knowledge in at least two of the following areas: machine learning (particularly neural networks), computer vision, visual odometry. - Experience in C++ and Python is mandatory - Experience in a using a popular machine learning frameworks such as Caffe or TensorFlow. - Strong problem solving skills - Students from outside of D-MAVT (particularly, from D-INFK, D-ITET, D-PHYS, and D-MATH) are also highly encouraged to apply.

Contact Details: If you are interested, please send your grade transcripts, CV, and a few sentences about your coding background to Igor Gilitschenski (igilitschenski@ethz.ch) and Thomas Schneider (thomas.schneider@mavt.ethz.ch).

SLAM Machine Learning Python C++ Robotics Computer Vision Convolutional Neural Networks (CNNs)


Labels: Semester project Master Thesis CLS Student Project (MPG ETH CLS)
Topics: Mathematical Sciences Information, Computing and Communication Sciences Behavioural and Cognitive Sciences

© 2017, Copyright Max-Planck-Gesellschaft - Imprint