An investigation of transfer learning for deep architectures in group activity recognition

Pervasive technologies permeating our immediate surroundings provide a wide variety of means for sensing and actuating in our environment, having a great potential to impact the way we live, but also how we work. In this paper, we address the problem of activity recognition in office environments, as a means for inferring contextual information in order to automatically and proactively assists people in their daily activities. To this end we employ state-of-the-art image processing techniques and evaluate their capabilities in a real-world setup. Traditional machine learning is characterized by instances where both the training and test data share the same distribution. When this is not the case, the performance of the learned model is deteriorated. However, often times, the data is expensive or difficult to collect and label. It is therefore important to develop techniques that are able to make the best possible use of existing data sets from related domains, relative to the target domain. To this end, we further investigate in this work transfer learning techniques in deep learning architectures for the task of activity recognition in office settings. We provide herein a solution model that attains a 94% accuracy under the right conditions.