VaLW: multi-centric motion forecasting with variable length window

Trajectory prediction plays a crucial role in achieving autonomous driving, as it significantly reduces driving risks by predicting the movement trajectory of other vehicles. The key challenge lies in effectively encoding scene information and generating accurate multimodal results for each agent. To address this challenge, we propose a graph neural network framework that enables multi-centric modeling of relationships between heterogeneous inputs. This framework establishes various spatiotemporal adjacency relationships among scene nodes, leveraging graph attention mechanisms to allow each scene node to learn neighborhood features effectively and generate scene context enriched with valuable information. To tackle the problem of declining prediction accuracy as the prediction time increases, we propose a variable length window structure. The structure consists of a long window prediction module for multi-agent multimodal prediction, followed by a short window optimization module for refining the predictions. By utilizing this structure, we successfully strike a balance between model size and prediction accuracy. To validate our proposed model, we conducted experiments on the Argverse 1 motion forecasting dataset, and the results showcased excellent predictive performance.