SCTP: Scene Compliant Trajectory Prediction Using Diffusion and Point Clouds
Authors
Andersen, Mads Bach ; Hansen, Simon Andreas
Term
4. term
Education
Publication year
2023
Abstract
Predicting how people and vehicles will move is vital for self-driving cars, but human behavior is unpredictable. Many systems use generative neural networks such as Conditional Variational Autoencoders (CVAEs) and Generative Adversarial Networks (GANs), which learn from data to produce plausible futures. These methods are strong, yet they can still struggle to produce natural, collision-free paths. We explore Denoising Diffusion Probabilistic Models (DDPMs), a newer kind of generative model that creates outputs by gradually removing noise. DDPMs have been successful in image, audio, and time-series tasks, but have seen limited use in trajectory prediction. We show that DDPMs can generate realistic, scene-compliant trajectories—that is, paths that respect the road layout and nearby obstacles—and can compete with leading methods through our model, SCTP. Because fast prediction is crucial in autonomous driving, we address the slow runtime of diffusion models by combining a leapfrogging sampling strategy with a point-cloud map representation (a set of points describing the environment). This makes inference more than twice as fast compared with traditional approaches. Finally, we show that closed-loop prediction—producing trajectories step by step while repeatedly consulting the map—can perform as well as common one-shot methods that predict the whole path at once. This repeated map evaluation helps the model avoid collisions.
At forudsige, hvordan mennesker og køretøjer bevæger sig, er afgørende for selvkørende biler, men menneskelig adfærd er uforudsigelig. Mange systemer bruger generative neurale netværk, fx Conditional Variational Autoencoders (CVAEs) og Generative Adversarial Networks (GANs), som lærer af data at foreslå mulige fremtidsforløb. Disse metoder er stærke, men kan stadig have svært ved at lave naturlige, kollisionsfri baner. Vi undersøger Denoising Diffusion Probabilistic Models (DDPM'er), en nyere type generativ model, der skaber resultater ved gradvist at fjerne støj. DDPM'er har haft succes i billeder, lyd og tidsserier, men er kun i begrænset omfang brugt til forudsigelse af bevægelsesbaner. Vi viser, at DDPM'er kan generere realistiske baner i overensstemmelse med omgivelserne—altså forløb, der respekterer vejens udformning og forhindringer—og kan konkurrere med de bedste metoder via vores model, SCTP. Da hurtige forudsigelser er vigtige i autonom kørsel, adresserer vi diffusion-modellers langsomme køretid ved at kombinere en leapfrogging-strategi med en punktsky-repræsentation af kortet (en mængde punkter, der beskriver omgivelserne). Det gør inferensen mere end dobbelt så hurtig sammenlignet med traditionelle tilgange. Endelig viser vi, at lukket-sløjfe-forudsigelser—hvor banen genereres trin for trin med gentagen vurdering af kortet—kan yde lige så godt som almindelige one-shot-metoder, der laver hele banen på én gang. Den gentagne vurdering hjælper modellen med at undgå kollisioner.
[This apstract has been rewritten with the help of AI based on the project's original abstract]