基於互動關係重要性篩選的路口車輛軌跡預測

Abstract

預測車輛未來軌跡是自動駕駛中的一項重要的任務。先前的研究主要集中在為每個車輛預測個別的未來軌跡。然而，在現實世界的情境中，尤其是在複雜的場景，如路口，更重要的是預測多個車輛未來的聯合軌跡。在這份研究裡，我們將問題範疇縮小，專注於僅為中心車輛預測不會和周圍其他車輛未來發生碰撞的軌跡。我們提出了一個名為PITP的框架，它利用初始的多車輛軌跡預測模型，為中心車輛的每條可能軌跡選擇不同需要關注的車輛。PITP利用這些被關注的車輛，獲取沿著軌跡附近區域的場景特徵。隨後，它進一步利用這些特徵通過交叉關注學習所有車輛和地圖元素之間的互動。我們的實驗在Argoverse2上進行，其中包括從中擷取的複雜路口場景。結果顯示，PITP在公開資料上取得了具有競爭性的表現。此外，和目前的邊緣預測模型相比，PITP在特別挑戰性的場景中表現出顯著的改進。

Predicting future trajectories of road agents is a crucial task in autonomous driving. Previous studies primarily concentrate on forecasting individual trajectories for each agent. Nevertheless, in real-world situations, particularly in complex scenarios such as intersections, prioritizing the generation of joint predictions for future trajectories across multiple agents becomes more crucial. In this work, we narrow down the problem to focus exclusively on predicting scene-compliant trajectories for a target agent. We propose a framework, termed PITP, which leverages the initial multi-agent prediction model to select different interactive agents for each potential trajectory of the target agent. PITP utilizes these interactive agents to acquire near local scene context features along each path. Subsequently, it further employs these features to learn interactions among all agents and map elements through cross attention. Our experiments were conducted on Argoverse2, with the complex intersection scenario extracted from it. The results indicate that PITP achieves competitive performance on benchmark and demonstrates significant improvement in challenging scenes compared to the current marginal prediction model.