Monocular visual odometry as part of the Robotics course at Cambridge University
Mars rovers require a high degree of autonomy as the long signal round trip time (> 6min) makes real-time control from Earth impossible. The Curiosity rover is equipped with several sensors that support its planning capabilities and uses mainly motor feedback to determine its location. However, the feedback from motors alone is not sufficient for precise localisation, e.g. wheels can slip. Pavol's task, as part of the Mobile Robot Systems course at Cambridge University, was to correct for sensor drifts of the Curiosity rover, using visual odometry (VO) and a camera feed.
The work uses the public imagery from Curiosity to perform monocular visual odometry, however, the image feeds are not continuous and data cleaning is needed. Pavol utilized Deepnote to create a new curated image dataset CuriosityVO suitable for visual odometry.
Data gathering, filtering, and visualisation, experimental reproducibility
Python, OpenCV, FTP Image Repositories
Python
Pavol Drotár is a MEng student of Computer Science at the University of Cambridge
Reproducibility
Pavol chose Deepnote for building the dataset due to the reproducible environment that can be shared with other researchers. Instead of sharing the entire dataset, a link to the deepnote notebook can be shared. This has two benefits: 1. not just the data, but the documentation about how it was retrieved can be presented in an environment that is available online to anyone, 2. the dataset can be kept up-to-date by re-running the notebook.
Dataset visualisation
Deepnote allows quick visualisations of data frames which is handy for examining the correctness and statistics of the dataset in between filtering steps.
Data gathering
Pavol utilised different sources of information, an FTP repository of images, and a JSON collection of image meta-data. Deepnote allows the interaction with both of these, and moreover, is flexible to modularly add additional data sources such as AWS buckets and SQL databases.
The fully reproducible environment in Deepnote was crucial to my research.
There is a great potential for Deepnote in the research community.
Outcome
Using the curated dataset built in Deepnote, Pavol was able to run a visual odometry pipeline on the martian terrain. A video demonstrating the odometry pipeline is available on YouTube. The work offered new insight into what are the challenging aspects of martian VO: surfaces covered with small rocks make feature matching challenging and cause drifts in position estimates.
Join the world's best data teams and get started with Deepnote
No credit card required. Run your first notebook in seconds.