Trajectory segmentation using greedy best rst change point detection based on trajectory similarity

Keywords

Loading...
Thumbnail Image

Issue Date

2021-06-25

Language

en

Document type

Journal Title

Journal ISSN

Volume Title

Publisher

Title

ISSN

Volume

Issue

Startpage

Endpage

DOI

Abstract

In day to day life nding structure in information is easy for humans, but this is not so easy for smart machine learning methods. They require structured data in order to be smart and become smarter, the more data, the better. There is a immense amount of data in existence, but most of this data is unstructured. Especially trajectory data, that consists of a series of x and y coordinates, is di cult to structure because of the large variety of information it can represent. Many methods exist that can nd structure in trajectory data by splitting it into many small segments, creating a trajectory segmentation. Many of these methods lack the ability to t to patterns of di erent length and thus struggle to nd segments of di erent lengths. Many methods also use complex approaches that are di cult to reason about and thus to generalize to other data, where a simple rule-based strategy could yield similar results. Most notably trajectory similarity methods are simple methods to compare trajectories. They are an ideal candidate to be used in the context of trajectory segmentation, but are scarcely being used. We identify a new solution that uses a trajectory similarity-based cost function and multiple temporal time scales to identify the points of largest pattern change, by tting windows to patterns in the data. This has the simple intuition that, if any two patterns are very di erent, they are likely to be different processes and should thus be separate segments. In a comprehensive test using numerous di erent parameter con guration on a dataset consisting of 5 trajectories of mice, we compare our method to two of the best change point detection methods: Binary segmentation and Pelt. The results indicate that our method achieves the best performance when the window size can be optimized for the data. When our method is not optimized for the data, Pelt generally has the better performance.

Description

Citation

Faculty

Faculteit der Sociale Wetenschappen