DsMCL: Dual-Level Stochastic Multiple Choice Learning for Multi-Modal Trajectory Prediction

TL;DR

This paper introduces DsMCL, a novel method for multi-modal vehicle trajectory prediction that does not require modal labels and achieves state-of-the-art results with a single forward pass.

Contribution

The paper proposes a dual-level stochastic multiple choice learning approach that captures multi-modal vehicle behaviors without explicit labels, improving prediction accuracy.

Findings

Significant improvement on NGSIM and HighD datasets.

Achieves state-of-the-art performance in multi-modal trajectory prediction.

Compatible with various existing prediction frameworks.

Abstract

For both driving safety and efficiency, automated vehicles should be able to predict the behavior of surrounding traffic participants in a complex dynamic environment. To accomplish such a task, trajectory prediction is the key. Although many researchers have been engaged in this topic, it is still challenging. One of the important and inherent factors is the multi-modality of vehicle motion. Because of the disparate driving behaviors under the same condition, the prediction of vehicle trajectory should also be multi-modal. At present, related researches have more or less shortcomings for multi-modal trajectory prediction, such as requiring explicit modal labels or multiple forward propagation caused by sampling. In this work, we focus on overcoming these issues by pointing out the dual-levels of multi-modal characteristics in vehicle motion and proposing the dual-level stochastic multiple choice learning method (named as DsMCL, for short). This method does not require modal labels and can implement a comprehensive probabilistic multi-modal trajectory prediction by a single forward propagation. By experiments on the NGSIM and HighD datasets, our method has proven significant improvement on several trajectory prediction frameworks and achieves state-of-the-art performance.