Impact of Visual Noise in Activity Recognition Using Deep Neural Networks - An Experimental Approach

Abstract

The popularity of deep learning methods has increased significantly, in no small part due to their impressive performance in several application scenarios. This paper focuses on recognising activities in an in-vehicle environment and measuring the impact that factors such as resolution, aspect ratio, field of view and framerate have on the performance of the model. The use of deep learning methodologies in recent years has increased the amount of data required to train and test the models. However, such data is often insufficient, unavailable, or lacks suitable properties. Publicly available action recognition datasets have been analysed, collected, and prepared to assess the classification results in such scenarios, which provides important guidance for use in a real-world setting.