[Paper Review] A survey on trajectory clustering analysis

Jiang Bian, Dayong Tian+2 more|arXiv (Cornell University)|Feb 20, 2018

Data Management and Algorithms96 references79 citations

TL;DR

This survey reviews trajectory clustering methods, categorizing them into unsupervised, supervised, and semi-supervised approaches, and discusses representations, distance measures, models, and future directions.

ABSTRACT

This paper comprehensively surveys the development of trajectory clustering. Considering the critical role of trajectory data mining in modern intelligent systems for surveillance security, abnormal behavior detection, crowd behavior analysis, and traffic control, trajectory clustering has attracted growing attention. Existing trajectory clustering methods can be grouped into three categories: unsupervised, supervised and semi-supervised algorithms. In spite of achieving a certain level of development, trajectory clustering is limited in its success by complex conditions such as application scenarios and data dimensions. This paper provides a holistic understanding and deep insight into trajectory clustering, and presents a comprehensive analysis of representative methods and promising future directions.

Motivation & Objective

• Provide a holistic understanding of trajectory clustering and its applications in surveillance, abnormal behavior detection, crowd analysis, and traffic control.
• Summarize data representations, feature extraction, and distance metrics used to measure trajectory similarity.
• Review representative methods across unsupervised, supervised, and semi-supervised categories and discuss their advantages and limitations.
• Highlight challenges such as data heterogeneity, variable trajectory lengths, and computational scalability, and propose future directions.

Proposed method

• categorize trajectory clustering methods into unsupervised, supervised, and semi-supervised. Discuss trajectory representations (unified length, transformation, resampling, sub-trajectories, POI, scale-invariant features). Compare distance measures (Euclidean, Hausdorff, Bhattacharyya, Frechet, DTW, LCSS, and others) and their trade-offs. Describe unsupervised models: densely clustering (DBSCAN, K-means/EM/FCM), hierarchical clustering, and spectral clustering. Describe supervised approaches: nearest-neighbor, SVM, Bayesian/inference models, neural networks (CNNs, DNNs, SOM). Outline semi-supervised methods and their typical workflows.
• Compare methodological categories in terms of computational considerations, strengths, and limitations (e.g., density-based vs. spectral vs. hierarchical approaches).
• Provide a synthesis of how trajectory length, representation, and distance metrics influence clustering outcomes.

Experimental results

Research questions

• RQ1 What are the main representations and distance measures used for trajectory clustering across different algorithmic categories?
• RQ2 How do unsupervised, supervised, and semi-supervised trajectory clustering methods compare in terms of performance, scalability, and applicability to various domains?
• RQ3 What are the current challenges and limitations in trajectory clustering, and what future directions are proposed?
• RQ4 How do preprocessing steps like transformation, resampling, and sub-trajectory extraction impact clustering results?
• RQ5 What roles do neural networks and probabilistic models play in trajectory clustering, and what are their trade-offs?

Key findings

• Trajectory clustering methods are organized into unsupervised, supervised, and semi-supervised categories with distinct data requirements and use cases.
• A variety of trajectory representations (unified length, curve fitting, sub-trajectories, POIs, scale-invariant features) and distance metrics (DTW, LCSS, Hausdorff, Frechet, Euclidean, Bhattacharyya) are used to measure similarity, each with different trade-offs.
• Densely clustering models (e.g., DBSCAN) dominate many trajectory analyses but can struggle with density variation; hierarchical and spectral clustering offer complementary strengths and limitations.
• Supervised approaches leverage labeled data to improve clustering and can use nearest-neighbor, SVM, Bayesian, and neural network-based methods to classify or segment trajectories.
• Semi-supervised methods balance labeled and unlabeled data to reduce labeling burden while maintaining performance.
• The survey discusses practical considerations such as trajectory length unification, computational complexity, and scene-specific adaptations, and points to future research directions.