VeTraSS: Vehicle Trajectory Similarity Search Through Graph Modeling and Representation Learning

TL;DR

VeTraSS addresses trajectory similarity search in autonomous driving by constructing tailored multi-scale graphs from trajectory data and learning embeddings with a multi-layer attention-based GNN. It transforms high-dimensional trajectory data into node embeddings via an end-to-end pipeline, enabling efficient similarity search. The method achieves state-of-the-art results on Porto and Geolife datasets under Fréchet and Hausdorff distances, outperforming both non-learning and prior learning-based approaches. This supports accurate trajectory analysis and safe autonomous navigation in real-world scenarios.

Abstract

Trajectory similarity search plays an essential role in autonomous driving, as it enables vehicles to analyze the information and characteristics of different trajectories to make informed decisions and navigate safely in dynamic environments. Existing work on the trajectory similarity search task primarily utilizes sequence-processing algorithms or Recurrent Neural Networks (RNNs), which suffer from the inevitable issues of complicated architecture and heavy training costs. Considering the intricate connections between trajectories, using Graph Neural Networks (GNNs) for data modeling is feasible. However, most methods directly use existing mathematical graph structures as the input instead of constructing specific graphs from certain vehicle trajectory data. This ignores such data's unique and dynamic characteristics. To bridge such a research gap, we propose VeTraSS -- an end-to-end pipeline for Vehicle Trajectory Similarity Search. Specifically, VeTraSS models the original trajectory data into multi-scale graphs, and generates comprehensive embeddings through a novel multi-layer attention-based GNN. The learned embeddings can be used for searching similar vehicle trajectories. Extensive experiments on the Porto and Geolife datasets demonstrate the effectiveness of VeTraSS, where our model outperforms existing work and reaches the state-of-the-art. This demonstrates the potential of VeTraSS for trajectory analysis and safe navigation in self-driving vehicles in the real world.

Figures (7)

• Figure 1: Trajectory similarity search pipeline. VeTraSS constructs the graph from the original vehicle trajectory data, followed by graph representation learning and embedding generation. The embedding is used for accurately searching alternative trajectories.
• Figure 2: Overview of the VeTraSS pipeline.Graph construction and embedding generation: The original high-dimensional trajectories are mapped into low-dimensional space for graph construction, where each node corresponds to each trajectory while edge connections represent trajectory similarity degrees. The multi-scale graph is input into a multi-scale GNN to generate accurate node embeddings. Trajectory similarity search: Given a query trajectory, VeTraSS generates the corresponding embedding vector and also finds the closest embedding vector to it. The closest embedding represents the most similar trajectory.
• Figure 3: Multi-scale graph construction. Edge between node $v_r$ and $v_p$ shows the similarity between trajectory $\mathcal{T}_r$ and $\mathcal{T}_p$. The nodes that are not connected in the $k^{th}$ layer (with a distance of 2) will form a connection in the $k+1^{th}$ layer, replacing the original edges.
• Figure 4: Graph representation learning module. The figure illustrates the $k^{th}$ layer of the model, where the inputs are the graph of layer $k$ ($G_k$) and the embedding of layer $k-1$ ($H^{k-1}$), and outputs the embedding of the $k^{th}$ layer ($H^{k}$). The final output of the model is the concatenation of each layer's output followed by an MLP module.
• Figure 5: Qualitative visualization.Left: Query trajectory. The three most similar trajectories are retrieved on the right. Right (a): Top-3 retrieval results of ground truth by computing the real distance as mentioned in Section \ref{['eval']}. Right (b): Top-3 retrieval results of VeTraSS. It is observed that the retrieval results of VeTraSS closely match the ground truth, qualitatively proving its effectiveness.