Exploring Shared Gaussian Occupancies for Tracking-Free, Scene-Centric Pedestrian Motion Prediction in Autonomous Driving

Engineering notes

To assess predictive performance, we compare our approach to state-of-the-art trajectory prediction methods, analyzing several metrics while keeping practical applications in mind. Evaluations on a dedicated pedestrian benchmark derived from the Argoverse 2 dataset demonstrate the model’s strong predictive accuracy and highlight the potential for tracking-free future developments.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

: This work introduces a scalable framework for pedestrian motion prediction in urban traffic, tailored for real-world applications in autonomous driving. Existing methods typically predict either individual objects, creating challenges with higher agent counts, or rely on discretized occupancy maps, sacrificing precision. To overcome these limitations, we propose a scene-centric transformer architecture with a cluster-based training approach, capturing pedestrian dynamics through combined probability distributions. This strategy enhances prediction efficiency as groups of nearby agents are unified into a shared representation, thus reducing computational load while still maintaining a continuous output format. Additionally, we investigate a tracking-free design, exploring the feasibility of accurate predictions based solely on object lists without explicit object association. To assess predictive performance, we compare our approach to state-of-the-art trajectory prediction methods, analyzing several metrics while keeping practical applications in mind. Evaluations on a dedicated pedestrian benchmark derived from the Argoverse 2 dataset demonstrate the model’s strong predictive accuracy and highlight the potential for tracking-free future developments.

Links and sources

• Official / arXiv page

Comments