Cross View Action Recognition

Cross View Action Recognition (CVAR) appraises a system's ability to recognise actions from viewpoints that are unfamiliar to the system. The state of the art methods that train on large amounts of training data rely on variation in the training data itself to increase their ability to tackle viewpoints changes. Therefore, these methods not only require a large scale dataset of appropriate classes for the application every time they train, but also correspondingly large amount of computation power for the training process leading to high costs, in terms of time, effort, funds and electrical energy. In this thesis, we propose a methodological pipeline that tackles change in viewpoint, training on small datasets and employing sustainable amounts of resources. Our method uses the optical flow input with a stream of a pre-trained model as-is to obtain a feature. Thereafter, this feature is used to train a custom designed classifier that promotes view-invariant properties. Our method only uses video information as input, in contrast to another set of methods that approach CVAR by using depth or pose input at the expense of increased sensor costs. We present a number of comparative analysis that aided the design of the pipelines, farther assessing the power of each component in the pipeline. The technique can also be adopted to existing, trained classifiers, with minimal fine-tuning, as this work demonstrates by comparing classifiers including shallow classifiers, deep pre-trained classifiers and our proposed classifier trained from scratch. Additionally, we present a set of qualitative results that promote our understanding of the relationship between viewpoints in the feature-space.