无人驾驶汽车周边车辆行为识别算法研究

Research on Behavior Recognition Algorithm of Surrounding Vehicles for Driverless Car

周边车辆行为识别对于提升无人驾驶汽车决策规划的合理性和控制安全性至关重要。传统的周边车辆行为识别方法识别精度普遍不高,且缺乏对交通主体相互之间邻域影响的考虑,算法鲁棒性较差。针对此,本文中提出了一种SLSTMAT(Social-LSTM-Attention)算法,创新性地引入目标车辆社交特征并通过卷积神经网络提取,建立了基于深度学习的车辆行为识别模型,应用注意力机制来捕捉行为时窗中的多时步信息,实现了周边车辆行为准确识别。采用HighD轨迹数据集和实车数据进行算法验证。结果表明,所提算法对周边车辆行为识别的准确率达94.01%,在目标车辆到达换道点的前1 s时刻行为识别精度达90%,具有较高的工程应用价值。

The behavior recognition of surrounding vehicles is very important to improve the decision-making rationality and the control safety for driverless cars.Traditional methods of surrounding vehicle’s behavior recognition suffer from low accuracy and poor robustness,without considering the interaction between adjacent traffic objects.In order to solve these problems,SLSTMAT(Social-LSTM-Attention)algorithm is proposed to achieve high accuracy for surrounding vehicle’s behavior recognition.The social characteristics of target vehicle are innovatively introduced and extracted by convolutional neural network.Based on deep learning,the recognition model for vehicle behavior is established.Attention mechanism is applied to capture multiple time-step information in the behavior time window.The algorithm is verified by HighD trajectory data set and real vehicle data.The results show that the accuracy rate of SLSTMAT for surrounding vehicle’s behavior recognition reaches 94.01%,and the precision of behavior recognition reaches 90%at 1s before the target vehicle driving to the lane change point,which has high engineering application value.