The ability to predict the intent of drivers from surrounding vehicles to change lanes is key to risk assessment and early danger warning systems. Since lane change trajectories are highly nonlinear, many studies have been performed on various machine learning algorithms using different features to predict a driver’s intent to change lanes. However, these algorithms use various features that cannot be obtained from the ego vehicle’s view point. In this paper, we define features that can be detected from the ego vehicle via on-board sensors and vehicle-to-vehicle communication (V2V). Gini Impurity is used to select the most appropriate features. Additionally, we compare several machine learning algorithms, including random forest (RF), support vector machine (SVM), long short term memory (LSTM), and gated recurrent unit (GRU), to find the best algorithm to predict lane changes. We evaluate the performance of these four algorithms on the Next Generation Simulation (NGSIM) dataset, which was collected by the Federal Highway Administration of the U.S. Department of Transportation. We use the I-80 dataset to train a lane changing prediction model and the US-101 dataset to test it. The test results indicate that RF had the best accuracy of the tested algorithms with an accuracy of 82 % in predicting lane changing intent.

能够预测周围车辆驾驶员改变车道的意图的能力是风险评估和早期危险预警系统的关键。由于换道轨迹是高度非线性的，因此已经对使用不同功能的各种机器学习算法进行了许多研究，以预测驾驶员换道的意图。但是，这些算法使用了从自我车辆的角度无法获得的各种功能。在本文中，我们定义了可以通过车载传感器和车对车通信（V2V）从自我车辆中检测到的特征。Gini不纯性用于选择最合适的功能。此外，我们比较了几种机器学习算法，包括随机森林（RF），支持向量机（SVM），长期短期记忆（LSTM）和门控循环单元（GRU），寻找最佳算法来预测车道变化。我们在下一代模拟（NGSIM）数据集上评估了这四种算法的性能，该数据集由美国交通部的联邦公路管理局收集。我们使用I-80数据集训练车道变更预测模型，并使用US-101数据集进行测试。测试结果表明，RF具有被测算法的最佳准确性，在预测车道变化意图方面的准确性为82％。