Use this URL to cite or link to this record in EThOS:
Title: Understanding egocentric human actions with temporal decision forests
Author: Garcia-Hernando, Guillermo
ISNI:       0000 0004 7655 3429
Awarding Body: University of London
Current Institution: Imperial College London
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Understanding human actions is a fundamental task in computer vision with a wide range of applications including pervasive health-care, robotics and game control. This thesis focuses on the problem of egocentric action recognition from RGB-D data, wherein the world is viewed through the eyes of the actor whose hands describe the actions. The main contributions of this work are its findings regarding egocentric actions as described by hands in two application scenarios and a proposal of a new technique that is based on temporal decision forests. The thesis first introduces a novel framework to recognise fingertip writing in mid-air in the context of human-computer interaction. This framework detects whether the user is writing and tracks the fingertip over time to generate spatio-temporal trajectories that are recognised by using a Hough forest variant that encourages temporal consistency in prediction. A problem with using such forest approach for action recognition is that the learning of temporal dynamics is limited to hand-crafted temporal features and temporal regression, which may break the temporal continuity and lead to inconsistent predictions. To overcome this limitation, the thesis proposes transition forests. Besides any temporal information that is encoded in the feature space, the forest automatically learns the temporal dynamics during training, and it is exploited in inference in an online and efficient manner achieving state-of-the-art results. The last contribution of this thesis is its introduction of the first RGB-D benchmark to allow for the study of egocentric hand-object actions with both hand and object pose annotations. This study conducts an extensive evaluation of different baselines, state-of-the art approaches and temporal decision forest models using colour, depth and hand pose features. Furthermore, it extends the transition forest model to incorporate data from different modalities and demonstrates the benefit of using hand pose features to recognise egocentric human actions. The thesis concludes by discussing and analysing the contributions and proposing a few ideas for future work.
Supervisor: Kim, Tae-Kyun Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral