The UGRacing society attempts to follow real-world organisation methods, splitting development areas into subteams each led by a subteam head, with a head software engineer deciding the long-term development workflow and a team principal to represent the team officially. Upon joining the team in 2021 in my first year of University, I was placed in the Perception subteam, which was responsible for maintaining and improving the systems which detect and represent the track during dynamic events.

When I initialy joined the team, the perception system was camera-based, and the team had only just acquired a new Velodyne VLP-16 LiDAR, but had written no software for it yet. My first project was to get the LiDAR sensor working. This took some time, as I had to dig through documentation and manuals not only to understand how such a sensor worked but also how the Robot Operating System (ROS) ecosystem worked – as it was the framework we used for writing our code.

Getting the sensor to produce a poitncloud was only half the battle, as it is not a trivial task to turn a cloud of thousands of points in 3-dimensional space into a representation of the track in front of the vehicle in an efficient way (such that it can be done ten times per second). As the track is delimited by cones, we are mainly interested in determining the 3d positions of these cones in front of the sensor. These cones are small in comparison to other objects which can be on the track and so a large part of any LiDAR object detection module is first noise removal / filtering. This can take on simple forms such as static cropping, and can be more sophisticated such as the use of dynamic ground plane identification and filtering – as the LiDAR will always return points representing the floor and constituting unusuful noise in our usecase.

After sucessfully developing a LiDAR perception system, my next task was to find a good way to combine estimates from both the LiDAR and Camera perception systems so that we could achieve a ‘best of both worlds’ scenario – where the fast update rate and good range of the camera could be combined with the higher precision and better field of view of the lidar to negotiate slow corners. This task required careful calibration between the camera and LiDAR, as their different mounting positions meant that cone observations were represented in different world frames. Additionally, data came in out of sync, so data had to be synchronised to be used and finally, observation noise can lead to ambiguity when attempting to correlate cone observation instances across sensors, as particularly noisy estimates can be misattributed to unrelated cones.

After completing this work and proving myself motivated and capable, I was chosen to be the new Head Software Engineer in the 2023-2024 season. During this season, my work continued to include software engineering, but also became a higher level management of the team, deciding upon software development philosophies, overarching goals and objectives and how our progress can be presented to the judges at the competition. During this year, I also embarked on the implementation of FastSLAM, as well as optimising our stereo camera perception. I achieved a significant speed up (50%) in our camera perception as well as significantly improving the robustness by making use of some high probability asusmptions to vastly simplify the 3d regression problem.