QCNet Aims to Better Guess Where Other Drivers Are Aiming to Boost the Safety of Autonomous Vehicles

Researchers from the City University of Hong Kong, the Hon Hai Research Institute, and Carnegie Mellon University have come up with a new approach to making autonomous vehicles better able to anticipate what other road users may do: query-centric trajectory projection.

"By integrating [our] technology into autonomous driving systems, the autonomous vehicles can effectively understand their surroundings, predict the future behavior of other users more accurately, and make safer and more human-like decisions, paving the way for safe autonomous driving," claims project lead Wang Jianping, professor in CityU's department of computer science, of the work's potential. "We plan to apply this technology to more applications in autonomous driving, including traffic simulations and human-like decision-making."

The technology in question: QCNet, a machine learning model for trajectory prediction — figuring out where other vehicles are going in order to avoid collisions. The system is built around a query-centric design ethos for scene encoding, boosting parallelism over rival approaches and using anchor-free queries followed by anchor-based queries in a refinement pass to offer higher performance and better-quality trajectory projections.

In simulation, the system shows promise: tested in Argoverses One and Two, a pair of open source driving data and map sets considered to be challenging benchmarks for autonomous vehicle testing, QCNet beat its rivals in both accuracy and performance — cutting online inference latency from 8ms to 1ms and boosting efficiency by over 85 per cent in the simulation's densest traffic scene, comprised of 190 separate road users across 169 map polygons.

The team's work has been published under open-access terms following its presentation at the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2023 (CVPR 2023); a copy, source code, and pre-trained models are available under the permissive Apache 2.0 license on the project's GitHub repository. The same team has also been working on an extended version dubbed QCNeXt; a preprint detailing this work is available on Cornell's arXiv server.