基于卡口上下文和深度置信网络的车辆轨迹预测模型研究

针对车辆轨迹预测中节点序列的时序特性和实际路网中的空间关联性,该文提出一种基于深度置信网络和SoftMax (DBN-SoftMax)轨迹预测方法.首先,考虑到轨迹在节点集合中的强稀疏性和一般特征学习方法对新特征的泛化能力不足,该文利用深度置信网络(DBN)较强的无监督特征学习能力,达到提取轨迹局部空间特性的目的;然后,针对轨迹的时序特性,该文采用逻辑回归的预测思路,用当前轨迹集在路网特征空间中的线性组合来预测轨迹;最后,结合自然语言处理领域中的词嵌入的思想,基于实际轨迹中节点存在的上下文关系,运用节点的向量集表征了节点间的交通时空关系.实验结果表明该模型不仅能够有效地提取轨迹特征,并且在拓扑结构复杂的路网中也能得到较好的预测结果.

Vehicle Trajectory Prediction Method Based on Intersection Context and Deep Belief Network

For the temporal features of trajectory intersection sequence and spatial correlation of the actual road network, a trajectory prediction method based on the Deep Belief Networks and SoftMax (DBN-SoftMax) is proposed. At first, considering the sparsity of trajectory in an intersection set and the insufficiency of generalization ability in general feature learning methods for new features, the strong unsupervised feature learning ability of Deep Belief Network (DBN) is used to extract the local spatial features of trajectory. Secondly, considering the temporal features of the trajectory, the logistic regression method and the linear combination of the current trajectory set in the road network features space are used to predict the trajectory. Finally, Based on the idea of word embedding in the field of natural language processing and the contextual relationship of intersections in the actual trajectory, the vector set of intersections is used to represent the spatiotemporal relationship of traffic between intersections. The experimental results show that the model can not only extract the trajectory features effectively, but also obtain better prediction performance in a road network with complex topology.