Unsupervised pattern and outlier detection for pedestrian trajectories using diffusion map #

Fanqi Zeng, Nikolai Bode, Thilo Gross, Martin Homer

11:30 Tuesday in 3Q16.

Part of the Collective behaviour and transport session.

Abstract #

Analyzing collective human behaviour, such as the evacuation of a crowd from a confined space, presents numerous challenges. Due to nonlinear dynamics, individual variation, and the sheer size of behavioural data, it can be difficult to efficiently study and understand these complex phenomena. In this work, we propose using diffusion map, an unsupervised manifold learning method, to uncover patterns in pedestrian behaviour. By defining a set of observables from the time series data of pedestrian trajectories, we construct a feature space. We examine two different types of crowd evacuation data: closed and open. The closed data is experimental, while the open data includes both experimental and simulated data. By applying the diffusion map to these feature spaces, we obtain the leading eigenvectors that contain valuable information about the collective behaviours. Our results highlight the leading variable that governs the dynamics of collective behaviours in the closed data, and allow us to differentiate between data sources in the open data. Additionally, we can identify anomalous individuals in the open data without the need to examine the original recordings. Our work provides an efficient approach to analyzing collective human behaviours without requiring extensive prior knowledge of the data.