Valeo.ai team, a member of the EXA4MIND consortium and leader of the project’s industrial application case, has developed a new autonomous driving method for predicting the future movements of nearby objects. It is called ‘Valeo4Cast: A Modular Approach to End-to-End Forecasting’, and is part of the methods or solutions that have participated in the AV2 2024 Unified Detection, Tracking, and Forecasting Challenge, organised by Argoverse.
Argoverse is one of the most popular automotive research datasets and since 2019 organises challenges to further improve open source databases. This year’s challenge, in which Valeo.ai presented its new method, aimed to improve on long-range detection (e.g. 100m – 150m) and non-linear motion forecasting. “Object detection and forecasting are fundamental components of embodied perception. These problems, however, are largely studied in isolation. We propose a joint detection, tracking, and multi-agent forecasting benchmark from sensor data. Although prior works have studied end-to-end perception, no large scale dataset or challenge exists to facilitate standardised evaluation for this problem”, explain on their website.
In autonomous driving, predicting the future movements of nearby objects, such as pedestrians, other cars and traffic lights, is crucial for safety. Valeo’s work focuses on motion prediction, which is anticipating where these objects will be based on their current and past movements. Traditionally, systems try to handle everything simultaneously, from detecting objects using sensors to predicting their future trajectories, a method known as ‘end-to-end’ forecasting. Instead, Valeo takes a modular approach by breaking the problem into smaller parts: detection (finding objects), tracking (following those objects over time) and forecasting (predicting their future movements). By improving each part separately and then tuning them together, they achieved better overall performance. This method far outperformed the usual end-to-end approach. With this seemingly simple, but totally innovative and demonstrably effective approach, ‘Valeo4Cast: A Modular Approach to End-to-End Forecasting’ is currently in 1st place on the leaderboard in the Motion Forecasting task and 2nd in the Tracking task.
‘Valeo4Cast: A Modular Approach to End-to-End Forecasting’, is the result of extensive and innovative research supported by EXA4MIND that will serve, among many other applications, to develop, demonstrate and improve the industrial application case for EXA4MIND.
If you want to stay updated about EXA4MIND, subscribe to our newsletter and follow us on X and LinkedIn!