CoLoss-GAN: Collision-Free Human Trajectory Generation with a Collision Loss and GAN

Humans navigate through crowds in a socially compliant manner and usually try to avoid close contact to strangers. An important skill that enables humans to achieve this is the accurate prediction of human movement. This skill is also desirable for autonomous mobile platforms for safe and socially conforming actions. Current research explores the challenging stochastic nature of a human's future trajectory using deep learning. However, it neglects the property that people move predominantly collision-free. We address this problem and introduce in this paper the CoLoss-GAN, a generative adversarial network (GAN) that encodes historical trajectories and generatively decodes future socially conforming trajectories. We propose an optimized state refinement and an effective pooling module, which learn feature representations that reliably encode human-to-human interactions and current human intentions. We encourage collision-freeness by formulating a crucial collision loss (CoLoss) that penalizes colliding predicted trajectories. Our experiments demonstrate performance on challenging real-world benchmarks, outperforming several state-of-the-art deterministic and generative baselines, in terms of accuracy and collision-freeness, with more plausible and nonlinear trajectory predictions.