SIFT-Motion Estimation (SIFT-ME): A New Feature for Human Activity Recognition

Action representation for robust human activity recognition is still a challenging problem. This thesis proposed a new feature for human activity recognition named SIFTMotion Estimation (SIFT-ME). SIFT-ME is derived from SIFT correspondences in a sequence of video frames and adds tracking information to describe human body motion. This feature is an extension of SIFT and is used to represent both translation and rotation in plane rotation for the key features. Compare with other features, SIFT-ME is new as it uses rotation of key features to describe action and it robust to the environment changes. Because SIFT-ME is derived from SIFT correspondences, it is invariant to noise, illumination, and small view angle change. It is also invariant to horizontal motion direction due to the embedded tracking information. For action recognition, we use Gaussian Mixture Model to learn motion patterns of several human actions (e.g., walking, running, turning, etc) described by SIFT-ME features. Then, we utilize the maximum log-likelihood criterion to classify actions. As a result, an average recognition rate of 96.6% was achieved using a dataset of 261 videos comprised of six actions performed by seven subjects. Multiple comparisons with existing implementations including optical flow, 2D SIFT and 3D SIFT were performed. The SIFT-ME approach outperforms the other approaches which demonstrate that SIFT-ME is a robust method for human activity recognition.