Human Action Recognition

Hongying Meng, Nick Pears, Zezhi Chen, Mike Freeman, Chris Bailey

Home | PhD applicants | Research | Publications | Google Scholar | Teaching | Bio

We have worked on human action recognition, with a view to embedding this functionality within the intelligent home.

Our approach has been to develop methods to detect and segment the moving person from the rest of the scene using adaptive background segmentation. We then use support vector machine (SVM) classifiers which operate on novel motion history representations.

Images below show adaptive foreground/backround segmentation with spatio-temporal filtering. Left is original image. On the right image, red is the desired foreground segmentation, blue is classified as a shadow.

In the image below, top-left shows original image, bottom-left shows desired segmentation as yellow pixels and shaow as green pixels. Bottom right shows segmentation with shadow removal, top right shows foreground superimposed on learnt background and hence shadow is removed from the original image.

The figure below shows a range of six different actions. After segmentation from the background, we store the motion history of these actions and decompose them according to the temporal patterns seen at each pixel. These patterns are used in a support vector machine framework to classify the motions into one of six classes.

Selected publications.

BACK to Nick Pears' Research Projects page.