On Motion Analysis in Computer Vision with Deep Learning: Selected Case Studies

Abstract

Motion analysis is one of the essential enabling technologies in computer vision. Despite recent significant advances, image-based motion analysis remains a very challenging problem. This challenge arises because the motion features are extracted directory from a sequence of images without any other meta data information. Extracting motion information (features) is inherently more difficult than in other computer vision disciplines.
In a traditional approach, the motion analysis is often formulated as an optimisation problem, with the motion model being hand-crafted to reflect our understanding of the problem domain. The critical element of these traditional methods is a prior assumption about the model of motion believed to represent a specific problem. Data analytics’ recent trend is to replace hand-crafted prior assumptions with a model learned directly from observational data with no, or very limited, prior assumptions about that model. Although known for a long time, these approaches, based on machine learning, have been shown competitive only very recently due to advances in the so-called deep learning methodologies.
This work's key aim has been to investigate novel approaches, utilising the deep learning methodologies, for motion analysis where the motion model is learned directly from observed data. These new approaches have focused on investigating the deep network architectures suitable for the effective extraction of spatiotemporal information. Due to the estimated motion parameters' volume and structure, it is frequently difficult or even impossible to obtain relevant ground truth data. Missing ground truth leads to choose the unsupervised learning methodologies which is usually represents challenging choice to utilize in already challenging high dimensional motion representation of the image sequence. The main challenge with unsupervised learning is to evaluate if the algorithm can learn the data model directly from the data only without any prior knowledge presented to the deep learning model during

In this project, an emphasis has been put on the unsupervised learning approaches. Owning to a broad spectrum of computer vision problems and applications related to motion analysis, the research reported in the thesis has focused on three specific motion analysis challenges and corresponding practical case studies. These include motion detection and recognition, as well as 2D and 3D motion field estimation.

Eyeblinks quantification has been used as a case study for the motion detection and recognition problem. The approach proposed for this problem consists of a novel network architecture processing weakly corresponded images in an action completion regime with learned spatiotemporal image features fused using cascaded recurrent networks.
The stereo-vision disparity estimation task has been selected as a case study for the 2D motion field estimation problem. The proposed method directly estimates occlusion maps using novel convolutional neural network architecture that is trained with a custom-designed loss function in an unsupervised manner.
The volumetric data registration task has been chosen as a case study for the 3D motion field estimation problem. The proposed solution is based on the 3D CNN, with a novel architecture featuring a Generative Adversarial Network used during training to improve network performance for unseen data.
All the proposed networks demonstrated a state-of-the-art performance compared to other corresponding methods reported in the literature on a number of assessment metrics. In particular, the proposed architecture for 3D motion field estimation has shown to outperform the previously reported manual expert-guided registration methodology.