Collective motion pattern learning for crowd segmentation and anomaly detection

Abstract

In scenes with collectively moving objects, to disregard the individual objects and take the entire group into consideration for motion characterization is a promising approach, and one of its important applications is crowd motion analysis. In crowd scenes, people's movement are mainly driven by different self-organized processes, and they usually move in groups. We consider each group as a single entity, and employ optical flow fields to represent people's movement in surveillance videos. In this thesis, we focus on extracting motion patterns of collectively moving objects from optical flow fields and apply the patterns to crowd segmentation and anomalous crowd activity recognition. Based on the observation that the collectively moving objects in a group have very similar motion behavior with their neighbors, we use the distribution of optical flow in a local region to represent the local motion pattern, and propose an approach for joint segmentation of collectively moving objects. Since a group of collectively moving objects correspond to a region where the optical flow field has high magnitude and high local direction coherence, our problem can be addressed by identifying these coherent optical flow field regions. The segmentation is performed through the minimization of a variational energy functional derived from the Bayes classification rule. In this model, we use a bag-of-words model to generate a codebook as a collection of prototypical optical flow patterns, and the class-conditional probability density functions for different regions are determined based on these patterns. The minimization of our proposed energy functional results in the gradient descent evolution of segmentation boundaries which are implicitly represented through level sets. The application of our proposed approach is to segment and track multiple groups of collectively moving objects in a large variety of real-world scenes. In addition, crowd movement in a local region can be approximated by a translational motion field based on the above observation. As a result, the local motion pattern can be represented as a translation domain parameter vector, and the problem of crowd segmentation is transformed into the problem of local-translational domain segmentation. We extend a translation domain segmentation model for unit vector fields to a local-translation domain segmentation (LTDS) model for general vector fields. In this model, an optical flow field is partitioned into local-translation domains characterizing different movement types of crowd regions, and the evolution of domain boundaries is derived from the Gateaux derivative of an objective functional. To represent crowd motion in an accurate and parsimonious way, we compute optical flow at the salient locations instead of at all the pixel locations, and further extend LTDS to the case of scattered motion field. Beside local motion patterns, we also study the global motion patterns of collective motion, and apply them to crowd activity recognition. We propose a Bayesian model to analyze crowd behavior and timely detect anomalous situations. Since people naturally escape from a place when unexpected events happen, we consider that crowd escape behavior is a possible indication of an unexpected event occurring. Specifically, we introduce the concepts of potential destinations and divergent centers to characterize crowd motion in the above two cases respectively, and the global motion patterns are represented by the corresponding distribution of optical flow. Although only data characterizing normal crowd behaviors are included in the training set, the density functions associated with the escape model can also be adaptively updated using newly observed data. In addition, the identified divergent centers indicate possible locations at which the unexpected events occur. Escape detection is finally performed based on the proposed Bayesian framework.

Date of Award	15 Feb 2013
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Hau San WONG (Supervisor)