| 基于LSTM网络的驾驶意图识别及车辆轨迹预测 |
| |
| 引用本文： | 季学武,费聪,何祥坤,刘玉龙,刘亚辉.基于LSTM网络的驾驶意图识别及车辆轨迹预测[J].中国公路学报,2019,32(6):34-42. |
| |
| 作者姓名： | 季学武 费聪 何祥坤 刘玉龙 刘亚辉 |
| |
| 作者单位： | 1. 清华大学 汽车安全与节能国家重点实验室, 北京 100084;2. 华为技术有限公司 诺亚方舟实验室, 北京 100085 |
| |
| 基金项目： | 国家自然科学基金中国汽车产业创新发展联合基金项目（U1664263）；国家自然科学基金项目（51875302）；国家重点研发计划项目（2018YFB1600501） |
| |
| 摘 要： | 自动驾驶汽车需具备预测周围车辆轨迹的能力，以便做出合理的决策规划，提高行驶安全性和乘坐舒适性。运用深度学习方法，设计了一种基于长短时记忆（LSTM）网络的驾驶意图识别及车辆轨迹预测模型，该模型由意图识别模块和轨迹输出模块组成。意图识别模块负责识别驾驶意图，其利用Softmax函数计算出驾驶意图分别为向左换道、直线行驶、向右换道的概率；轨迹输出模块由编码器-解码器结构和混合密度网络（MDN）层组成，其中的编码器将历史轨迹信息编码为上下文向量，解码器结合上下文向量和已识别的驾驶意图信息预测未来轨迹；引入MDN层的目的是利用概率分布来表示车辆未来位置，而非仅仅预测一条确定的轨迹，以提高预测结果的可靠性和模型的鲁棒性。此外，将被预测车辆及其周围车辆组成的整体视为研究对象，使模型能够理解车-车间的交互式行为，响应交通环境的变化，动态地预测车辆位置。使用基于真实路况信息的NGSIM（Next Generation SIMulation）数据集对模型进行训练、验证与测试。研究结果表明：与传统的基于模型的方法相比，基于LSTM网络的轨迹预测方法在预测长时域轨迹上具有明显的优势，考虑交互式信息的意图识别模块具备更高的预判性和准确率，且基于意图识别的轨迹预测能降低预测轨迹与真实轨迹间的均方根误差，显著提高轨迹预测精度。
|
| 关 键 词： | 汽车工程 轨迹预测 意图识别 LSTM 交互式行为 |
| 收稿时间： | 2019-03-26 |
Intention Recognition and Trajectory Prediction for Vehicles Using LSTM Network |
| |
| JI Xue-wu,FEI Cong,HE Xiang-kun,LIU Yu-long,LIU Ya-hui.Intention Recognition and Trajectory Prediction for Vehicles Using LSTM Network[J].China Journal of Highway and Transport,2019,32(6):34-42. |
| |
| Authors: | JI Xue-wu FEI Cong HE Xiang-kun LIU Yu-long LIU Ya-hui |
| |
| Affiliation: | 1. State Key Laboratory of Automotive Safety and Energy, Tsinghua University, Beijing 100084, China;2. Noah's Ark Lab, Huawei Technologies, Beijing 100085, China |
| |
| Abstract: | Autonomous vehicles often need to predict the trajectories of surrounding vehicles for planning and decision making. In this paper, a model for intention recognition and trajectory prediction based on long short-term memory (LSTM) network is proposed. The proposed model comprises an intention recognition module and a trajectory output module. The intention recognition module was employed for identifying the driving intention. The Softmax function was incorporated in the intention recognition module for calculating the probabilities of left lane change, lane-keeping, and right lane change. An encoder-decoder structure and a mixture density network (MDN) layer were included in the trajectory output module. The encoder converted the past trajectory information into the context vector. Subsequently, the decoder combined the context vector and the intention recognition information for predicting future trajectories. The MDN layer was employed for representing the future position of a vehicle with its probability distribution rather than with a particular trajectory, which improved the reliability of prediction results and the robustness of the proposed model. Additionally, the compositions of a predicted vehicle and its surroundings were both taken into account, which aided the proposed model in analyzing the interactions among vehicles. Hence, the proposed model can dynamically predict vehicle trajectories according to variations in traffic conditions. The NGSIM data set based on the information of actual road conditions was employed for training, validating, and testing the proposed model. Experimental results indicate that the proposed method based on LSTM network has several advantages over conventional model-based methods with respect to trajectory prediction, especially in a long prediction horizon. Interactive information can ensure that the intention recognition module has high anticipative ability and accuracy. Furthermore, trajectory prediction based on intention recognition can significantly reduce the root-mean-square errors in predicted trajectories with respect to the ground truth, thereby leading to significant improvement in trajectory prediction accuracy. |
| |
| Keywords: | automotive engineering trajectory prediction intention recognition long short-term memory interaction behavior |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《中国公路学报》浏览原始摘要信息 |
|
点击此处可从《中国公路学报》下载全文 |
|
