Abstract: To ensure safe and efficient interactions between advanced autonomous vehicles and pedestrians,rapid and accurate recognition of pedestrian crossing intentions has become a focal point of research for scholars worldwide.The current state of research in pedestrian crossing intention recognition was reviewed focusing on three key areas:pedestrian characteristics,contextual information and related model applications.In terms of pedestrian characteristics,the study was categorized into skeletal posture tracking,pedestrian visual tracking and pedestrian trajectory prediction,with a detailed examination of the progression from 2D to 3D human joint posture techniques.The discussion of contextual information addressed the integration of environmental context,visual cues and pedestrian trajectory fusion for intention recognition.Regarding model applications,the utilization of deep learning models including CNNs,LSTMs,GNNs and Transformers were analyzed.At the relevant technical level,five key technologies related to pedestrian crossing intention were summarized.In summary,the main achievements and limitations of existing research were summarized,the current challenges were identified and the future research directions to enhance the safety of interactions between pedestrians and autonomous vehicles were proposed.

Key words: autonomous driving, pedestrian crossing intention recognition, human-machine interaction, review

摘要： 为确保自动驾驶汽车与过路行人之间的交互安全和效率,快速有效地识别行人过路意图成为国内外学者的研究热点。本文从行人特征、情景信息、相关模型和相关技术4个方面分析了当前的研究现状。在行人特征层面将其划分为骨架姿势跟踪、行人视觉追踪、行人轨迹预测,总结了人体关节姿势识别技术从2D过渡到3D的演变过程。从上下文环境信息、视觉线索融合、行人轨迹融合阐述行人意图识别的情景信息。在相关模型应用层面,主要分析了卷积神经网络（CNN）、长短期记忆网络（ LSTM）、生成对抗网络（GAN）、图卷积神经网络（GCN）、自注意力的神经网络（Transformer）5类深度学习模型。在相关技术方面,总结了5种有关行人过路意图识别的关键技术。本文总结了现有研究的主要成果和不足,提出了当前存在的问题和未来研究方向,以期提高过路行人与自动驾驶汽车交互的安全性。

关键词: 自动驾驶, 行人过路意图识别, 人机交互, 综述