GDTS: Goal-Guided Diffusion Model with Tree Sampling for Multi-Modal Pedestrian Trajectory Prediction

TL;DR

GDTS, a novel Goal-Guided Diffusion Model with Tree Sampling for multi-modal trajectory prediction that leverages goal estimation to guide the generation of the diffusion network and achieves comparable state-of-the-art performance with real-time inference speed in public datasets.

Abstract

Accurate prediction of pedestrian trajectories is crucial for improving the safety of autonomous driving. However, this task is generally nontrivial due to the inherent stochasticity of human motion, which naturally requires the predictor to generate multi-modal prediction. Previous works leverage various generative methods, such as GAN and VAE, for pedestrian trajectory prediction. Nevertheless, these methods may suffer from mode collapse and relatively low-quality results. The denoising diffusion probabilistic model (DDPM) has recently been applied to trajectory prediction due to its simple training process and powerful reconstruction ability. However, current diffusion-based methods do not fully utilize input information and usually require many denoising iterations that lead to a long inference time or an additional network for initialization. To address these challenges and facilitate the use of diffusion models in multi-modal trajectory prediction, we propose GDTS, a novel Goal-Guided Diffusion Model with Tree Sampling for multi-modal trajectory prediction. Considering the "goal-driven" characteristics of human motion, GDTS leverages goal estimation to guide the generation of the diffusion network. A two-stage tree sampling algorithm is presented, which leverages common features to reduce the inference time and improve accuracy for multi-modal prediction. Experimental results demonstrate that our proposed framework achieves comparable state-of-the-art performance with real-time inference speed in public datasets.

Figures (3)

• Figure 1: Comparison of standard sampling algorithm and our proposed tree sampling algorithm for multi-modal prediction. Since the trunk stage of our algorithm only needs to run once for multiple predictions, the total number of diffusion steps is fewer than the standard sampling algorithm, therefore accelerating the inference speed.
• Figure 2: The architecture of our GDTS framework. GDTS consists of a goal estimation module and a diffusion-based trajectory prediction module. The goal estimation module predicts the goal distribution to estimate multiple goals, and the feature of history motion augmented with estimated goals information are then fed to the trajectory prediction module. The trajectory prediction module uses the two-stage tree sampling algorithm to denoise the trajectory from Gaussian noise $\widetilde{Y}^k$ to $\widetilde{Y}^{k-1}$ iteratively, conditioned on the inputted feature. During tree sampling, the common feature first acts as the guidance of denoising to generate a general initialization for further denoising in the trunk stage. Then in the branch stage, the diverse features "guide" the general initialization into different modalities to predict multi-modal future trajectories.
• Figure 3: Visualization of history trajectory and future prediction of three agents in different scenes on the ETH/UCY and SDD. The observed trajectory is in red, the goal estimations obtained from the goal estimation module are in blue, the final predictions are in cyan, and the ground truth goal and trajectory are in yellow. The goals are highlighted as star points.