Paper ThAT1.5
Liu, Haochen (Nanyang Technological University), Huang, Zhiyu (Nanyang Technological University), Mo, Xiaoyu (Nanyang Techonological University), Lv, Chen (Nanyang Technological University)
Multi-Modal Motion Prediction with Group-Wise Modal Assignment Transformer for Autonomous Driving
Scheduled for presentation during the Invited Session "Learning-powered and Knowledge-driven Autonomous Driving I" (ThAT1), Thursday, September 26, 2024,
11:50−12:10, Salon 1
2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada
This information is tentative and subject to change. Compiled on December 26, 2024
|
|
Keywords Automated Vehicle Operation, Motion Planning, Navigation, Sensing, Vision, and Perception, Network Modeling
Abstract
Forecasting accurate and consistent future states of traffic participants within the knowledge of uncertainty is imperative for enhancing the social synergies and driving safety in autonomous driving system, particularly under interactive scenarios. Tackling multi-modality through direct estimation or regressions remains the major challenge, primarily due to the need to balance the granularity of uncertainties with the sparsity of positive training samples to be updated for respective decoding modes. This work introduces GTR, a multi-modal motion prediction framework with a group-wise modal allocation scheme for Transformer-enabled trajectory decoding. Our approach includes several key steps to tackle this challenge. Firstly, we introduced the group-wise allocation strategy, a plugged-in decoding initialization for each modality, thereby densely increase training diversity of positive queries for respective modality. Additionally, a miss-rate optimization pipeline is instantiated which further maximize the discriminated margins for positive decoding queries. This neat decoding strategy achieved compelling prediction accuracy, social consistency, and outstanding performance across primary metrics in the Waymo Open Motion Dataset (WOMD) leaderboard.
|
|