Prediction of Lane-Changing Maneuvers with Automatic Labeling and Deep Learning
Prediction of Lane-Changing Maneuvers with Automatic Labeling and Deep Learning
复制标题
DOI：
10.1177/0361198120922210
复制
发表时间：
2020-06-12
影响因子：
1.7
通讯作者：
Antoniou, Constantinos
中科院分区：
文献类型：
--
作者：
Mahajan, Vishal;Katrakazas, Christos;Antoniou, Constantinos
近年来，公路安全引起了人们极大的研究兴趣，特别是随着互联和自动驾驶汽车（cav）等创新技术迅速成为现实。识别和预测驾驶意图是避免碰撞的基础，因为它可以为驾驶员及其附近的车辆提供有用的信息。然而，最先进的机动预测需要使用大型标记数据集，这需要大量的处理，并可能阻碍实时应用。在本文中，开发并提出了一个端到端机器学习模型，用于使用有限数量的特征从未标记数据中预测变道机动。该模型建立在一个全新的综合数据集（即highD）上，该数据集由配备摄像头的无人机从德国高速公路上获得。采用基于密度的聚类来识别变道和保持车道的机动，然后训练支持向量机（SVM）模型来学习聚类标签的边界，并自动标记新的原始数据。然后将标记的数据输入到长短期记忆（LSTM）模型中，该模型用于预测机动类。分类结果表明，该方法可以有效地实时预测车道变化，平均检测时间至少为3 s，虚警率很小。利用未标记数据和车辆特征作为特征增加了该方法的可转移性及其在公路安全方面的实际应用前景。
Highway safety has attracted significant research interest in recent years, especially as innovative technologies such as connected and autonomous vehicles (CAVs) are fast becoming a reality. Identification and prediction of driving intention are fundamental for avoiding collisions as it can provide useful information to drivers and vehicles in their vicinity. However, the state-of-the-art in maneuver prediction requires the utilization of large labeled datasets, which demand a significant amount of processing and might hinder real-time applications. In this paper, an end-to-end machine learning model for predicting lane-change maneuvers from unlabeled data using a limited number of features is developed and presented. The model is built on a novel comprehensive dataset (i.e., highD) obtained from German highways with camera-equipped drones. Density-based clustering is used to identify lane-changing and lane-keeping maneuvers and a support vector machine (SVM) model is then trained to learn the boundaries of the clustered labels and automatically label the new raw data. The labeled data are then input to a long short-term memory (LSTM) model which is used to predict maneuver class. The classification results show that lane changes can efficiently be predicted in real-time, with an average detection time of at least 3 s with a small percentage of false alarms. The utilization of unlabeled data and vehicle characteristics as features increases the prospects of transferability of the approach and its practical application for highway safety.