C$^{2}$INet: Realizing Incremental Trajectory Prediction with Prior-Aware Continual Causal Intervention

TL;DR

Comparative evaluations on three real and synthetic complex datasets against state-of-the-art methods demonstrate that the proposed Continual Causal Intervention method consistently achieves reliable prediction performance, effectively mitigating confounding factors unique to different scenarios.

Abstract

Trajectory prediction for multi-agents in complex scenarios is crucial for applications like autonomous driving. However, existing methods often overlook environmental biases, which leads to poor generalization. Additionally, hardware constraints limit the use of large-scale data across environments, and continual learning settings exacerbate the challenge of catastrophic forgetting. To address these issues, we propose the Continual Causal Intervention (C$^{2}$INet) method for generalizable multi-agent trajectory prediction within a continual learning framework. Using variational inference, we align environment-related prior with posterior estimator of confounding factors in the latent space, thereby intervening in causal correlations that affect trajectory representation. Furthermore, we store optimal variational priors across various scenarios using a memory queue, ensuring continuous debiasing during incremental task training. The proposed C$^{2}$INet enhances adaptability to diverse tasks while preserving previous task information to prevent catastrophic forgetting. It also incorporates pruning strategies to mitigate overfitting. Comparative evaluations on three real and synthetic complex datasets against state-of-the-art methods demonstrate that our proposed method consistently achieves reliable prediction performance, effectively mitigating confounding factors unique to different scenarios. This highlights the practical value of our method for real-world applications.

Figures (9)

• Figure 1: Left: In real-world applications, various environmental scenarios often contain confounding variables (denoted as $C$), such as regulations, customs, and road conditions, that influence trajectory data. To address this, our method constructs a causal model designed to mitigate the effects of spurious correlations on prediction outcomes. For instance, predicting a turn at an intersection and a lane change on a highway may be conflated, as both trajectories exhibit an initial directional shift. Right: The proposed C$^{2}$INet model incorporates a causal intervention trajectory prediction framework and a Continual Memory module with a prior queue. Utilizing a min-max training strategy, the model is optimized while acquiring optimal continual prior for newly added scenarios.
• Figure 2: The ADE variations of the validation sets across five task scenarios and their average performance during the training process on the ETH-UCY dataset. The x-axis represents the number of completed training epochs, while the y-axis denotes the corresponding metric values.
• Figure 3: Visualization of the posterior distribution for different models with 2D latent space.
• Figure 4: The Average Displacement Error (ADE) on the ETH-UCY dataset for each task, averaged over five runs under continual learning settings. The x-axis represents the sequence of completed tasks, while the y-axis indicates the corresponding metric values.
• Figure 5: The Average Displacement Error (ADE) on the Synthesis dataset for each task, averaged over five runs under continual learning settings. The x-axis represents the sequence of completed tasks, while the y-axis indicates the corresponding metric values.