After the publication of the Perception record and completing his Object recognition competition 2019, Lyft shared a new one today Body – The prediction record – contains the logs of the movements of cars, pedestrians and other obstacles encountered by the fleet of 23 autonomous vehicles in Palo Alto. At the same time, the company plans to introduce one Challenge This will instruct the participants to predict the movement of traffic agents.
A longstanding research problem in the self-driving area has been to create models that are robust and reliable enough to predict traffic movements. Lyft's data set focuses on motion prediction by taking into account the motion of the types of traffic that the fleet has used to cross, such as cars, cyclists and pedestrians. This movement is derived from data collected by the sensor suite mounted on the roof of Lyft's vehicles, which, for example, captures lidar and radar values when the vehicles travel tens of thousands of kilometers:
- Logs of over 1,000 hours of traffic agent movement
- 170,000 scenes, each lasting about 25 seconds
- 16,000 miles of data from public roads
- 15,000 semantic card notes
- The underlying semantic HD map of the area
Combined with the semantic high-resolution maps created by Lyft's teams in Palo Alto, London and Munich, the body contains the parts needed to create predictive models that allow vehicles to choose safe trajectories in certain scenarios, says Sacha Arnoud and Peter Ondruska from Lyft. "Data is the fuel for experimenting with the latest machine learning techniques, and limited access to high-quality, self-driving quality data shouldn't hinder experimentation with this research problem," they wrote in a blog post. "With this data set and competition we want to strengthen the research community, accelerate innovation and give insights into the problems to be solved from the perspective of a mature program (for autonomous vehicles)."
The training data set and the associated Python-based software kit are available today. Test and validation kits will be released as part of the competition, which starts on Google's Kaggle platform in August and brings in a total of $ 30,000 in prizes.
"We believe that self-driving will be an integral part of a more accessible, safer and more sustainable transport system," added Arnoud and Ondruska. “By sharing data with the research community, we hope to shed light on important and unsolved challenges in self-driving. Together we can realize the benefits of self-driving earlier. "
Lyft's second challenge kicked off months after Waymo expanded its public driving data set and launched the Waymo Open Dataset competition, worth $ 110,000. Were winners announced In mid-June during a workshop at the 2020 conference on Computer Vision and Pattern Recognition (CVPR), which took place online this year due to the coronavirus pandemic.