融合意图推断与时空交互建模的行人轨迹预测方法

近年来，行人轨迹预测技术快速发展并在智能运载装备领域得以初步应用。然而，现有行人轨迹预测技术仍存在诸多不足，如信息利用单一、难以建模行人与环境的时空耦合交互关系、难以实现长短时轨迹综合预测，严重制约了该技术在智能运载装备中的实际应用。为此，本项目研究融合意图推断与时空交互建模的行人轨迹预测方法，具体内容包括：(1) 融合多源信息的行人意图推断。建立考虑输入误差及丢失的推断模型，充分利用各个传感器的优势信息，实现高可靠性意图推断；(2) 行人时空注意力机制研究与交互建模。构建行人时空耦合交互模型，在时空非一致性背景下探究行人时空注意力机制，实现行人与环境交互关系精确建模；(3) 基于长短时域多模型融合的行人轨迹预测。结合行人意图推断和交互建模方法，探究长短时域模型融合机制，构建多时域轨迹预测模型，最终开展方法测试验证。本项目将为行人轨迹预测在智能运载装备中的应用提供重要的理论方法支撑。

In recent years, pedestrian trajectory prediction technology has developed rapidly and has been initially applied in the field of intelligent transportation equipment. However, the existing pedestrian trajectory prediction technology still has many defects, e.g., it just took single information as input, it lacked the ability to model the space-time coupling interaction between pedestrians and the environment, and it cannot implement comprehensive long- and short-term trajectory prediction, which severely restricts the practical application in intelligent transportation equipment. Therefore, this project researches a pedestrian trajectory prediction method combining intention inference and attention mechanism, including: (1) pedestrian intention inference based on the fusion of multi-source information: we establish firstly an inference model that takes into account input errors and losses, and then make full use of the superior information of each sensor to achieve high-reliability inference intent; (2) pedestrian spatiotemporal attention mechanism research and interactive modeling, we construct a pedestrian spatiotemporal coupling interaction model, which can explore the pedestrian spatiotemporal attention mechanism under the background of spatiotemporal inconsistency; (3) pedestrian trajectory prediction based on the fusion of multiple models in long and short time domains: we firstly combine the pedestrian intention inference and interactive modeling methods to explore the fusion mechanism of long and short time domain models. Based on these, a multi-time domain trajectory prediction model is constructed, and the experiment tests and verification are carried out finally. This project will provide important a significant theoretical and methodological support for the application of pedestrian trajectory prediction in intelligent and connected transportation equipment.