A Survey on Learning-Based Trajectory Similarity Computation

Trajectory Similarity Computation (TSC) quantifies the similarity or dissimilarity between trajectories, which is a fundamental task in trajectory data analysis. However, the high computational cost of traditional trajectory distance measures has become a key bottleneck for large-scale trajectory data analysis. With advancements in deep learning, learning-based TSC methods map trajectories into a uniform embedding space, enabling constant-time similarity computations and improving efficiency. Additionally, these methods exhibit robustness to low-quality data and adapt to various similarity measures through diverse learning strategies. This paper surveys recent learning-based TSC methods, classifying them by representation form (sequence, image, and graph) and learning strategy(supervised and self-supervised). We analyze their principles and characteristics, focusing on intra-trajectory representation learning, inter-trajectory embedding learning, and training data construction. We also review publicly available datasets and evaluation metrics, and discuss current challenges and future research directions in TSC.