Why
This dataset started as an idea to use a neural network for lidar cone detection instead of the regular geometric methods. It was theorised that I would become increasingly hard to increase both accuracy and range of detections without requiring longer execution times derived from larger pointclouds. As we started experimenting with lidar temporal scan agregation we realised that geometric methods would not scale well. We also saw the opportunity of having much improved detection precision and less false positives using neural networks since they would have the ability to infer from shape, intensity and perhaps even temporal information.
As I (Bernardo) started working on my Masters Thesis in Temporally Aggregated LiDAR Object Detection for autonomous street driving, I saw further potential of these networks to have a low computation time and more stable performance. More interesting it became clear how these networks can learn from context. The idea of the network being able to learn how a Formula Student track should look like and how cone positions relate to each other to form a cohesivew line, filtering out things that would otherwise be indistinguishable from cones was very appealing.
Who we are
We are a group of Aluminis and students from the Chalmers Formula Student. Having worked and been founders of the driverless system at Chalmers Formula Student we kept trying to contribute to the team and as we saw pootential with this project we decided to try it out and publish our results to the entire formula student community. We saw hwo much work a project like this would take and how it was quite hard to justify for the core team to do it.
Bernardo Taveira
Head of Autonomous 2022/2023
Currently an alumni of Chalmers Formula Student (2021-2023) & FST Lisboa (2018-2021)
“I said this was the last year 5 years ago… but I think I am starting to get the hang of it.”