SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model

TL;DR

This paper proposes SingularTrajectory, a diffusion-based universal trajectory prediction framework to reduce the performance gap across the five tasks, and adopts a diffusion-based predictor to further enhance the prototype paths using a cascaded denoising process.

Abstract

There are five types of trajectory prediction tasks: deterministic, stochastic, domain adaptation, momentary observation, and few-shot. These associated tasks are defined by various factors, such as the length of input paths, data split and pre-processing methods. Interestingly, even though they commonly take sequential coordinates of observations as input and infer future paths in the same coordinates as output, designing specialized architectures for each task is still necessary. For the other task, generality issues can lead to sub-optimal performances. In this paper, we propose SingularTrajectory, a diffusion-based universal trajectory prediction framework to reduce the performance gap across the five tasks. The core of SingularTrajectory is to unify a variety of human dynamics representations on the associated tasks. To do this, we first build a Singular space to project all types of motion patterns from each task into one embedding space. We next propose an adaptive anchor working in the Singular space. Unlike traditional fixed anchor methods that sometimes yield unacceptable paths, our adaptive anchor enables correct anchors, which are put into a wrong location, based on a traversability map. Finally, we adopt a diffusion-based predictor to further enhance the prototype paths using a cascaded denoising process. Our unified framework ensures the generality across various benchmark settings such as input modality, and trajectory lengths. Extensive experiments on five public benchmarks demonstrate that SingularTrajectory substantially outperforms existing models, highlighting its effectiveness in estimating general dynamics of human movements. Code is publicly available at https://github.com/inhwanbae/SingularTrajectory .

Figures (6)

• Figure 1: An overview of our SingularTrajectory framework. All relevant human trajectory prediction tasks can be represented in our Singular space, a unified feature embedding space for human dynamics. Using the embedding features, our diffusion-based model for a universal trajectory prediction makes prediction for all the tasks in this same space.
• Figure 2: Visualization of the trajectories in Singular space. (a) The circle and triangle markers indicate the history and future trajectory coordinates, respectively. Each color also refers to each associated task. (b-d) Each marker in the Singular space corresponds to each trajectory in raw data. And, the slicing planes, representing a set of arrows, mean human dynamics of straight forward motions and turning motions.
• Figure 3: An example of the adaptive anchor generation. (a) The initial prototype anchor $\mathcal{P}$ is placed on the last observation coordinate of a person. In this instance, four prototype paths (highlighted in red) are incorrectly placed at the non-traversable locations. (b) Vector field $\vec{F}_{I}(x,y)$ is computed to guide toward in the nearest traversable areas. (c) The initial prototype paths are then tailored to the environment using the vector field.
• Figure 4: Visualization of prediction results on (a) momentary observation task and (b) few-shot task. To aid visualization, the best trajectory among $S\!=\!20$ samples are reported.
• Figure 5: Visualization of prediction consistency across five tasks. The more consistent the prediction is the better.