Causal Intervention for Human Trajectory Prediction with Cross Attention Mechanism

Human trajectory Prediction (HTP) in complex social environments plays a crucial and fundamental role artificial intelligence systems. Conventional methods make use of both history behaviors interactions to forecast future trajectories. However, we demonstrate that the environment is confounder misleads model learn spurious correlations between To end this, first formulate environment, variables into structural causal analyze causalities among them. Based on intervention rather than conventional likelihood, propose Social Environment ADjustment (SEAD) method, remove confounding effect environment. The core our method implemented by Cross Attention (SCA) module, which universal, simple effective. Our has consistent improvements ETH-UCY datasets with three baseline models achieves competitive performances existing methods.