基于机器学习与双向时间序列数据的车辆行驶行为预测模型

Vehicle Driving Behavior Prediction Model Based on Machine Learning and Bi-Directional Time Series Data

摘要: 自驾车辆跟驰时需密切关注前车的行驶状态，以此采取加减速、制动等驾驶决策。现有的行为预测模型未考虑与前导车的耦合关系，且持续感知具有一定局限。因此本文提出基于机器学习与双向时间序列数据的车辆行驶行为预测模型。该模型以轻量方式感知前车，且引入优化的注意力机制，解决交通流密集时跟车频繁出现遮挡的误检问题。提取出二维行为向量以双向序列形式输入双层预测模型中，更精确预测跟驰行为数据变化趋势。实验结果表明，提出的模型感知精度达84.56%、召回率达80.37%，相比基准模型平均每帧耗时降低4.95 ms，并且对前导车的行为预测准确率较高，能为自驾车辆跟驰提供有效的行驶决策。

Abstract: Self-driving vehicles need to pay close attention to the driving status of the preceding vehicle when following, so as to take driving decisions such as acceleration, deceleration and braking. The exist-ing behavior prediction model does not consider the coupling relationship with the leading vehicle, and the continuous perception has certain limitations. Therefore, this paper proposes a prediction model of vehicle driving behavior based on machine learning and bi-directional time series data. The model senses the leading vehicle in a lightweight manner and introduces an optimized atten-tion mechanism to solve the false detection problem of frequent occlusion of following vehicles when the traffic flow is dense. The two-dimensional behavior vectors are extracted and input into the two-layer prediction model in the form of two-way sequences to predict the change trend of fol-lowing behavior data more accurately. The experimental results show that the proposed model has 84.56% perception accuracy and 80.37% recall rate, and the average time per frame is reduced by 4.95 ms compared with the benchmark model, and the prediction accuracy of the leading vehicle behavior is higher, which can provide effective driving decision for following vehicles.