- When: 27th February 2018 14:00 - 15:00
- Where: Cole 1.33a
- Series: School Seminar Series
- Format: Seminar
This talk describes work in our research group using computer vision along with other sensor modalities to recognise (i) actions in which people manipulate objects, and (ii) social interactions and their participants.
Activities such as those involved in food preparation involve interactions between hands, tools and manipulated objects that affect them in visually complex ways making recognition of their constituent actions challenging. One approach is to represent properties of local visual features with respect to trajectories of tracked objects. We explore an example in which reference trajectories are provided by visually tracking embedded inertial sensors. Additionally, we propose a vision method using discriminative spatio-temporal superpixel groups, obtaining state-of-the-art results (compared with published results using deep neural networks) whilst employing a compact, interpretable representation.
Continuous analysis of social interactions from wearable sensor data streams has a range of potential applications in domains including healthcare and assistive technology. I will present our recent work on (i) detection of focused social interactions using visual and audio cues, and (ii) identification of interaction partners using face matching. By modifying the output activation function of a deep convolutional neural network during training, we obtain an improved representation for open-set face recognition.
Prof. Stephen McKenna co-leads the Computer Vision and Image Processing (CVIP) group at the University of Dundee where he is Chair of Computer Vision and Computing’s Head of Research. His interests lie primarily in biomedical image analysis, computer vision, and applied machine learning.