Trajectory Prediction in the Complex Junction Scene Using Information Fusion and Behavior Pattern Recognition

Predicting trajectories in the context of complex environments is a task that goes beyond its initial perception. While it may appear as a straightforward sequence generation problem, there are additional factors that must be considered, such as road information and the interactions with other vehicles within real-world scenes. To tackle this challenge, deep learning techniques have been employed. By leveraging neural network models, it becomes possible to integrate diverse sensory signals and map information, thereby constructing comprehensive models that account for the heterogeneity of world states. Consequently, these models enable the inference of highly multimodal distributions for future trajectories. This paper proposes a future prediction model based on information fusion and behavioral pattern recognition. For the surrounding pedestrians and vehicles, we first judge their behavior patterns. Then try to use the target agent history data, other agent history data, and roadmap data, and merge these elements. The proposed model is evaluated using the Waymo Open Motion Dataset, and the experimental results demonstrate its superiority over most existing mainstream models in terms of metrics such as minADE and minFDE. These results highlight the significant improvement in accuracy and validity achieved by the proposed model in trajectory prediction.