Unsupervised human activity analysis for intelligent mobile robots
Declaration The candidate confirms that the work submitted is his/her own, except where work which has formed part of a jointly authored publication has been included. The contribution of the candidate and the other authors to this work has been explicitly indicated below. The candidate confirms that appropriate credit has been given within the thesis where reference has been made to the work of others. QSRLib: a software library for online acquisition of qualitative spatial relations from
... wledgements I gratefully acknowledge the financial support provided by the EU STRANDS project (Spatio-Temporal Representations and Activities for Cognitive Control in Long-Term Scenarios) and funded by the European Community's Seventh Framework Programme, Cognitive Systems and Robotics, project reference 600623. I would like to thank colleagues in the School of Computing at the University of Leeds and in the STRANDS project consortium (http://strands-project.eu) for their valuable input and direction. Finally, I would especially like to thank my supervisors Professors Anthony Cohn and David Hogg for their support. Abstract The success of intelligent mobile robots in daily living environments depends on their ability to understand human movements and behaviours. One goal of recent research is to understand human activities performed in real human environments from long term observation. We consider a human activity to be a temporally dynamic configuration of a person interacting with key objects within the environment that provide some functionality. This can be a motion trajectory made of a sequence of 2-dimensional points representing a person's position, as well as more detailed sequences of high-dimensional body poses, a collection of 3-dimensional points representing body joints positions, as estimated from the point of view of the robot. The limited field of view of the robot, restricted by the limitations of its sensory modalities, poses the challenge of understanding human activities from obscured, incomplete and noisy observations. As an embedded system it also has perceptual limitations which restrict the resolution of the human activity representations it can hope to achieve. In this thesis an approach for unsupervised learning of activities implemented on an autonomous mobile robot is presented. This research makes the following novel contributions: 1) A qualitative spatial-temporal vector space encoding of human activities as observed by an autonomous mobile robot. 2) Methods for learning a low dimensional representation of common and repeated patterns from multiple encoded visual observations. In order to handle the perceptual challenges, multiple abstractions are applied to the robot's perception data. The human observations are first encoded using a leg-detector, an upper-body image classifier, and a convolutional neural network for pose estimation, while objects within the environment are automatically segmented from a 3-dimensional point cloud representation. Central to the success of the presented framework is mapping these encodings into an abstract qualitative space in order to generalise patterns invariant to exact quantitative positions within the real world. This is performed using a number of qualitative spatial-temporal representations which capture different aspects of the relations between the human subject and the objects in the environment. The framework auto-generates a vocabulary of discrete spatial-temporal descriptors extracted from the video sequences and each observation is represented as a vector over this vocabulary. Analogously to information retrieval on text corpora we use generative probabilistic techniques to recover latent, semantically meaningful, concepts in the encoded observations in an unsupervised manner. The relatively small number of concepts discovered are defined as multinomial distributions over the vocabulary and considered as human activity classes, granting the robot a high-level understanding of visually observed complex scenes. We validate the framework using, 1) A dataset collected from a physical robot autonomously patrolling and performing tasks in an office environment during a six week deployment, and 2) a high-dimensional "full body pose" dataset captured over multiple days by a mobile robot observing a kitchen area of an office environment from multiple view points. We show that the emergent categories from our framework align well with how humans interpret behaviours and simple activities. Our presented framework models each extended observation as a probabilistic mixture over the learned activities, meaning it can learn human activity models even when embedded in continuous video sequences without the need for manual temporal segmentation, which can be time consuming and costly. Finally, we present methods for learning such human activity models in an incremental and continuous setting using variational inference methods to update the activity distribution online. This allows the mobile robot to efficiently learn and update its models of human activity over time, discarding the raw data, allowing for life-long learning.