Considering Cyclists' Aggressiveness and Bounded Rationality: A Self-learning Motion Prediction Model for Cyclists

Deep learning models have been widely used in trajectory prediction. However, they typically focus on developing a pre-training model by learning average features from trajectory data and are difficult to learn and predict heterogeneous behavior patterns. Especially in the case of cyclists, the prediction of low-probability risky behavior is a major challenge for autonomous driving systems. Considering that cyclists are likely to express more subjectivity than vehicle drivers, this paper proposes a self-learning motion prediction framework for adaptively predicting cycling behavior based on cyclists' psychological hidden variables. Instead of the time series modeling of traditional approaches, this framework views it as a dynamic estimation problem based on candidate trajectories. The aggressiveness and rationality levels of cyclists are set as the self-learning parameters because of their significant impact on cycling decisions. In addition, the feature identification methods and online learning strategies corresponding to the above features are proposed for self-learning. The performance of the proposed model has been verified using the bicycle trajectory data at an intersection in Shanghai, China. Compared with the state-of-the-art and baseline models, the self-learning model shows lower prediction errors. And it provides a significant improvement in complex scenarios. Then, an individual example is given to demonstrate how self-learning works. Finally, the paper visualizes and analyzes the characteristics of the psychological variables among cyclists.