[论文解读] Social-BiGAT: Multimodal Trajectory Forecasting using Bicycle-GAN and Graph Attention Networks

Vineet Kosaraju, Amir Sadeghian+4 more|arXiv (Cornell University)|Jul 4, 2019

Autonomous Vehicle Technology and Safety被引用 277

一句话总结

Social-BiGAT 使用图注意力网络和 Bicycle-GAN 启发的潜在编码来生成多模态、社会和物理上可行的行人轨迹预测，在标准基准上优于先前方法。

ABSTRACT

Predicting the future trajectories of multiple interacting agents in a scene has become an increasingly important problem for many different applications ranging from control of autonomous vehicles and social robots to security and surveillance. This problem is compounded by the presence of social interactions between humans and their physical interactions with the scene. While the existing literature has explored some of these cues, they mainly ignored the multimodal nature of each human's future trajectory. In this paper, we present Social-BiGAT, a graph-based generative adversarial network that generates realistic, multimodal trajectory predictions by better modelling the social interactions of pedestrians in a scene. Our method is based on a graph attention network (GAT) that learns reliable feature representations that encode the social interactions between humans in the scene, and a recurrent encoder-decoder architecture that is trained adversarially to predict, based on the features, the humans' paths. We explicitly account for the multimodal nature of the prediction problem by forming a reversible transformation between each scene and its latent noise vector, as in Bicycle-GAN. We show that our framework achieves state-of-the-art performance comparing it to several baselines on existing trajectory forecasting benchmarks.

研究动机与目标

• Motivate accurate and multimodal pedestrian trajectory forecasting for autonomous systems and social robots.
• Model rich social interactions and scene context to improve prediction realism.
• Introduce a bidirectional mapping between trajectories and latent noise to capture multimodality.
• Incorporate physical scene cues via attention mechanisms to enhance generalization.
• Evaluate against established baselines on standard trajectory datasets.

提出的方法

• Represent pedestrians as nodes in a fully connected graph and apply Graph Attention Networks to learn social interactions.
• Encode past pedestrian trajectories and scene context into latent features.
• Use a Bicycle-GAN inspired latent encoder to create a bijection between noise and trajectories for multimodal output.
• Condition a decoder LSTM on concatenated pedestrian, social, and physical context features plus latent noise to generate future trajectories.
• Train with dual discriminators (local-pedestrian and global-scene) to enforce realism at multiple scales.
• Optimize with adversarial losses, reconstruction loss for noise (Lz), trajectory reconstruction loss (Ltraj), and KL divergence to match Gaussian latent distribution.

实验结果

研究问题

• RQ1 Can Social-BiGAT learn multimodal distributions of pedestrian trajectories while modeling global social interactions?
• RQ2 Does the combination of graph attention for social cues and Bicycle-GAN style latent encoding improve multimodal trajectory generation over prior GAN-based approaches?
• RQ3 How do local and global discriminators influence trajectory realism and diversity?
• RQ4 Does incorporating scene context via soft attention improve prediction accuracy across diverse scenes?

主要发现

• Social-BiGAT achieves the best performance among tested models, reducing average final displacement error (FDE) by 0.15 m on average compared to the prior state-of-the-art.
• Incorporating Graph Attention Networks (GAT) improves performance over baselines that do not model global social interactions.
• The combination of GAT with the latent encoder (BiGAN) yields the strongest results, highlighting multimodal generation benefits.
• Latent space modeling provides more robust predictions at lower sample counts (lower K), reducing the increase in ADE/FDE variability.
• Qualitative results show Social-BiGAT produces lower-variance, more realistic trajectories in crowded and collision-avoidance scenarios.