Abstract
One essential task for autonomous driving is to encode the information of a driving scene into vector representations so that the downstream task such as trajectory prediction could perform well. The driving scene is complicated, and there exists heterogeneity within elements, where they own diverse types of information i.e., agent dynamics, map routing, road lines, etc. Meanwhile, there also exist relativity across elements - meaning they have spatial relations with each other; such relations should be canonically represented regarding the relative measurements since the absolute value of the coordinate is meaningless. Taking these two observations into consideration, we propose a novel backbone, namely Heterogeneous Driving Graph Transformer (HDGT), which models the driving scene as a heterogeneous graph with different types of nodes and edges. For graph construction, each node represents either an agent or a road element and each edge represents their semantics relations such as Pedestrian-To-Crosswalk, Lane-To-Left-Lane. As for spatial relation encoding, instead of setting a fixed global reference, the coordinate information of the node as well as its in-edges is transformed to the local node-centric coordinate system. For the aggregation module in the graph neural network (GNN), we adopt the transformer structure in a hierarchical way to fit the heterogeneous nature of inputs. Experimental results show that the proposed method achieves new state-of-the-art on INTERACTION Prediction Challenge and Waymo Open Motion Challenge, in which we rank 1st and 2nd respectively regarding the minADE/minFDE metric.
Abstract (translated)
URL
https://arxiv.org/abs/2205.09753
