An integrated framework for driving risk evaluation that combines lane-changing detection and an attention-based prediction model

In recent years, the increase in traffic accidents has emerged as a significant social issue that poses a serious threat to public safety. The objective of this study is to predict risky driving scenarios to improve road safety.
On the basis of data collected from naturalistic driving real-vehicle experiments, a comprehensive framework for identifying and analyzing risky driving scenarios, which combines an integrated lane-changing detection model and an attention-based long short-term memory (LSTM) prediction model, is proposed. The performance of the 4 machine learning methods on the CULane data set is compared in terms of model running time and running speed as evaluation metrics, and the ultrafast network with the best performance is selected as the method for lane line detection. We compared the performance of LSTM and attention-based LSTM on the basis of the prediction accuracy, recall, precision, and F1 value and selected the better model (attention-based LSTM) for risky scenario prediction. Furthermore, Shapley additive explanation analysis (SHAP) is used to understand and interpret the prediction results of the model.
In terms of algorithm efficiency, the running time of the ultrafast lane detection network only requires 4.1 ms, and the average detection speed reaches 131 fps. For prediction performance, the accuracy rate of attention-based LSTM reaches 96%, the precision rate is 98%, the recall rate is 96%, and the F1 value is 97%.
The improved attention-based LSTM model is significantly better than the LSTM model in terms of convergence speed and prediction accuracy and can accurately identify risky scenarios that occur during driving. The importance of factors varies by risky scenario. The characteristics of the yaw rate, speed stability, vehicle speed, acceleration, and lane change significantly influence the driving risk, among which lane change has the greatest impact. This study can provide real-time risky scenario prediction, warnings, and scientific decision guidance for drivers.

近年来，交通事故频发已成为威胁公共安全的重大社会问题。本研究旨在通过预测危险驾驶场景以提升道路交通安全水平。
本研究基于自然驾驶实车实验采集的数据，提出了一套融合集成式车道变换检测模型与基于注意力机制的长短期记忆网络（Long Short-Term Memory, LSTM）预测模型的危险驾驶场景识别与分析综合框架。
本研究以模型运行时长与运行速度为评估指标，对比了4种机器学习方法在CULane数据集上的表现，选取综合性能最优的超快网络作为车道线检测方案。基于预测准确率、召回率、精确率与F1值，对比了LSTM与基于注意力机制的LSTM的模型性能，并选取更优的基于注意力机制的LSTM模型用于危险场景预测。此外，本研究采用夏普利可加性解释分析（Shapley Additive exPlanations, SHAP）对模型的预测结果进行理解与阐释。
在算法效率方面，超快车道检测网络的单帧运行时长仅为4.1毫秒，平均检测速度可达131帧每秒。在预测性能上，基于注意力机制的LSTM的准确率达96%，精确率为98%，召回率为96%，F1值为97%。
改进后的基于注意力机制的LSTM模型在收敛速度与预测精度上均显著优于传统LSTM模型，可精准识别驾驶过程中出现的危险场景。不同危险场景下的影响因素重要性存在差异：横摆角速度、车速稳定性、车辆速度、加速度以及车道变换行为对驾驶风险具有显著影响，其中车道变换行为的影响程度最大。本研究可为驾驶员提供实时危险场景预测、预警以及科学的决策指导。