Multi-agent trajectory prediction has been crucial for the safe decision-making of autonomous driving in complex transportation scenarios. Making accurate and real-time predictions is of great significance, and interaction modeling has shown great potential in improving prediction accuracy, but it is a non-trivial task to capture the complicated interaction mechanisms between agents. Existing learning-based methods that consider interaction modeling often have large model sizes, sacrificing the inference speed in pursuit of superior accuracy, making them hard for practical implementations. To tackle this issue, we present a lightweight Lane-Guided Vector Transformer (LGVT) to balance the tradeoff between the inference speed and model accuracy. Specifically, we employ the Attention Free Transformer module that eliminates the need for dot product self-attention to efficiently learn the spatial-temporal dependency and leverages a motion encoder to encode agent motion features. In addition, we use a lane-guided representation that defines a region centered along the ground truth trajectory to align the agent's motion states with lane constraints, which helps filter out the irrelevant lane segments and accelerate the learning process. We conduct extensive experiments on the Argoverse 1 dataset and compare our model with several baselines. Results indicate that our model can outperform the baseline HiVT-64 regarding the prediction accuracy metrics minADE, minFDE, and MR. We also achieve an approximate accuracy of the large version HiVT-128 with a 70% decrease in model parameters and a 20% increase in the inference speed.
A Lightweight Lane-Guided Vector Transformer for Multi-Agent Trajectory Prediction in Autonomous Driving
2024-09-24
2840350 byte
Conference paper
Electronic Resource
English