Recognition of Pedestrians’ Street-Crossing Intentions Based on Skeleton Features

Abstract

An integrated method is proposed to solve the problem of frequent conflicts between autonomous vehicles and pedestrians in the street crossing scene. The method involves pedestrian detection, tracking, and intention recognition. First, an enhanced YOLOv8 is introduced by combining the C2f_CA module to achieve accurate pedestrian detection, tracking and pose estimation. Second, a variety of intention recognition features are proposed to characterize the position and pose of pedestrians in spatial and time domains. Finally, by taking the feature data as input for the base learners, the intention classification model is proposed based on the Stacking model with SVM, KNN and random forest as the base learners and XGBoost as the meta learner. The experimental results show that the enhanced YOLOv8 improves the detection accuracy by 5.4% compared with the original model, and the intention recognition based on the Stacking model can achieve 94.0% accuracy on the JAAD dataset, which is improved by more than 3.4% compared with the existing intention recognition models. Furthermore, when different parts of a pedestrian are occluded, the accuracy of the Stacking model still reaches 65.8%–73.3%, which verifies the robustness of the proposed model. The proposed model provides reliable inputs for decision planning of autonomous vehicles, which is conducive to improving the safety of self-driving.

摘 要

针对过街场景下智能车与行人冲突多发的情况,提出了一套针对行人检测、跟踪和意图识 别的集成方法。首先提出基于C2f_CA 模块改进YOLOv8 模型完成对行人的准确检测、跟踪和姿态 估计;然后提出多种意图识别特征在空间和时域关系下表征行人的位置与姿态;最后以特征数据为 输入,基于以SVM、 KNN 和随机森林三者为基模型的Stacking 异质集成方法完成行人的意图识别 建模。对上述模型进行实验验证,结果表明,改进后的YOLOv8 模型相较于原模型检测精度提高了 5.4%,基于Stacking 异质集成模型的行为意图识别在JAAD 数据集上可以达到94.0%的准确率,相 比于现有的意图识别模型提升了3.4%以上;在行人不同部位被遮挡的情况下,模型的准确率依旧达 到65.8%∼73.3%,验证了该方法的鲁棒性。该方法为自动驾驶汽车的决策规划提供了可靠的输入, 有利于提高自动驾驶的安全性。