Matthew Kent Myers
Using Convolutional Neural Networks for motion classification.
- Email: [email protected]
Supervisors
- School of Engineering
- School of Computing
Project description
Traditional deep learning approaches to computer vision include Convolutional Neural Networks (CNNs). They have been very successful at image classification, object detection and semantic segmentation. But these traditional approaches are only capable of processing still images. They fail to incorporate vital motion data encoded into continuous video frames. Thus, they are incapable of carrying out the more complex task of motion and action classification within video data. In this study, we will incorporate temporal information into the CNN architecture. This follows on from progress made within the field in recent years.
We will produce a novel approach to human action localisation and classification. We will investigate and analyse the spatial and temporal features extracted by 3D CNN. Our analysis will allow for individual tracking within a video sequence. It will provide a novel, quantitative and nuanced comparison between similar motions within the video.
These extracted features will also allow for comprehensive comparisons and exploration between 3D CNN architecture and other motion incorporation approaches, such as optical flow.
My PhD is linked to the manufacturing industry via the sponsor company. Models and findings within the PhD will be applied to real world manufacturing data. We will explore various application issues and difficulties.
Interests
Machine learning, deep learning, artificial intelligence, computer vision.
Qualifications
- MPhys 1st Class, University of Sheffield