Trajectory Semanticization: A Method of Accompany Vehicle Discovery Inspired by Semantic Similarity

Abstract

The study of accompanying vehicles is a hot topic in the field of intelligent transportation. Because of the multiple selectivity of the traffic path and the loss of sampling in traditional companion vehicles discovery, the method based on path similarity mining will result in the omission of the companion candidates. This paper recognizes that the upstream and downstream relevance of trajectory intentions in traffic is similar to the contextual relevance of text semantics, as inspired by the semantic similarity of texts. Simultaneously, taking into account the generalization and tolerance of semantic processing, a companion vehicle discovery method based on “trajectory semantics” similarity is proposed. First, this paper proposes a trajectory semantic vectorized representation method trajectory semantic to vector (TS2vec), which realizes the low‐dimensional dense vectorization of the trajectory in the context of dynamic time slicing of the trajectory, fusion of the temporal and spatial characteristics of the trajectory, and text information. Then, based on the “trajectory pair,” this paper proposes the trajectory pair bidirectional GRU (TPBi‐GRU) model. This paper constructs forward and reverse subnetworks using the trajectory pair set—which is made up of the actual trajectory and ts sampled trajectories—realizes parameter transfer and contribution during training; gains a thorough understanding of trajectory semantics; and mines the internal relationship between vehicles more effectively. Finally, given the difference in the degree of contribution of the road shape in forming the adjoint pattern, and the sensitivity of the attention mechanism to local features, the attention mechanism is used to weigh the key nodes that affect the trajectory shape in order to obtain a more accurate trajectory representation. The experimental results show that the method in this paper can discover local and overall concomitant patterns more effectively and effectively overcome the interference of multiple selectivity of traffic paths on concomitant pattern mining.