Advances in wearable technologies are facilitating the understanding of human activities using first-person vision (FPV) for a wide range of assistive applications. In this thesis, we propose robust multiple motion features for human activity recognition from first-person videos. The proposed features encode discriminant characteristics from magnitude, direction and dynamics of motion estimated using optical flow. Moreover, we design novel virtual-inertial features from video, without using the actual inertial sensor, from the movement of intensity centroid across frames. Results on multiple datasets demonstrate that centroid-based inertial features improve the recognition performance of grid-based features. Moreover, we propose a multi-layer modelling framework that encodes hierarchical and temporal relationships among activities. The first layer operates on groups of features that effectively encode motion dynamics and temporal variations of intra-frame appearance descriptors of activities with a hierarchical topology. The second layer exploits the temporal context by weighting the outputs of the hierarchy during modelling. In addition, a post-decoding smoothing technique utilises decisions on past samples based on the confidence of the current sample. We validate the proposed framework with several classifiers, and the temporal modelling is shown to improve recognition performance. We also investigate the use of deep networks to simplify the feature engineering from firstperson videos. We propose a stacking of spectrograms to represent short-term global motions that contains a frequency-time representation of multiple motion components. This enables us to apply 2D convolutions to extract/learn motion features. We employ long short-term memory recurrent network to encode long-term temporal dependency among activities. Furthermore, we apply cross-domain knowledge transfer between inertial-based and vision-based approaches for egocentric activity recognition. We propose sparsity weighted combination of information from different motion modalities and/or streams. Results show that the proposed approach performs competitively with existing deep frameworks, moreover, with reduced complexity.

Recent advances in camera technology have made it possible to build a comfortable, wearable system which can capture the scene in front of the user throughout the day. Products based on this technology, such as GoPro and Google Glass, have generated substantial interest. In this thesis, I present my work on egocentric vision, which leverages wearable camera technology and provides a new line of attack on classical computer vision problems such as object categorization and activity recognition. The dominant paradigm for object and activity recognition over the last decade has been based on using the web. In this paradigm, in order to learn a model for an object category like coffee jar, various images of that object type are fetched from the web (e.g. through Google image search), features are extracted and then classifiers are learned. This paradigm has led to great advances in the field and has produced state-of-the-art results for object recognition. However, it has two main shortcomings: a) objects on the web appear in isolation and they miss the context of daily usage; and b) web data does not represent what we see every day. In this thesis, I demonstrate that egocentric vision can address these limitations as an alternative paradigm. I will demonstrate that contextual cues and the actions of a user can be exploited in an egocentric vision system to learn models of objects under very weak supervision. In addition, I will show that measurements of a subject's gaze during object manipulation tasks can provide novel feature representations to support activity recognition. Moving beyond surface-level categorization, I will showcase a method for automatically discovering object state changes during actions, and an approach to building descriptive models of social interactions between groups of individuals. These new capabilities for egocentric video analysis will enable new applications in life logging, elder care, human-robot interaction, developmental screening, augmented reality and social media.