Deep Multi Temporal Scale Networks for Human Motion Analysis

The movement of human beings appears to respond to a complex motor system that contains signals at different hierarchical levels. For example, an action such as ``grasping a glass on a table'' represents a high-level action, but to perform this task, the body needs several motor inputs that include the activation of different joints of the body (shoulder, arm, hand, fingers, etc.). Each of these different joints/muscles have a different size, responsiveness, and precision with a complex non-linearly stratified temporal dimension where every muscle has its temporal scale. Parts such as the fingers responds much faster to brain input than more voluminous body parts such as the shoulder. The cooperation we have when we perform an action produces smooth, effective, and expressive movement in a complex multiple temporal scale cognitive task. Following this layered structure, the human body can be described as a kinematic tree, consisting of joints connected. Although it is nowadays well known that human movement and its perception are characterised by multiple temporal scales, very few works in the literature are focused on studying this particular property. In this thesis, we will focus on the analysis of human movement using data-driven techniques. In particular, we will focus on the non-verbal aspects of human movement, with an emphasis on full-body movements. The data-driven methods can interpret the information in the data by searching for rules, associations or patterns that can represent the relationships between input (e.g. the human action acquired with sensors) and output (e.g. the type of action performed). Furthermore, these models may represent a new research frontier as they can analyse large masses of data and focus on aspects that even an expert user might miss. The literature on data-driven models proposes two families of methods that can process time series and human movement. The first family, called shallow models, extract features from the time series that can help the learning algorithm find associations in the data. These features are identified and designed by domain experts who can identify the best ones for the problem faced. On the other hand, the second family avoids this phase of extraction by the human expert since the models themselves can identify the best set of features to optimise the learning of the model. In this thesis, we will provide a method that can apply the multi-temporal scales property of the human motion domain to deep learning models, the only data-driven models that can be extended to handle this property. We will ask ourselves two questions: what happens if we apply knowledge about how human movements are performed to deep learning models? Can this knowledge improve current automatic recognition standards? In order to prove the validity of our study, we collected data and tested our hypothesis in specially designed experiments. Results support both the proposal and the need for the use of deep multi-scale models as a tool to better understand human movement and its multiple time-scale nature.