Marrying the covariance-matrix representation with a recently developed classification framework based on sparse linear representation, related to compressive sampling, has lead to improved action recognition accuracy. Some of the tunnel shape measurements used in the 13-dimensional feature vector

Combined with the nearest neighbor classification, this early approach produced very promising results.

The tunnel’s shape, that captures action dynamics, is then described by a covariance matrix computed from 13-dimensional feature vectors. In one of the early variants of the approach, background subtraction is applied to all video frames in order to extract moving object’s silhouette tunnel. Central to the developed framework is the use of covariance-matrix representation to characterize actions in a video, and leading to a family of algorithms based on different features and classification methods. Furthermore, the proposed algorithms have low storage and computational requirements thus making them suitable for real-time implementation. Summary: In this project, a new class of action recognition algorithms has been developed with performance on par or better than that of of state-of-the-art methods. There are many application areas that could greatly benefit from a fast, accurate, and robust solution: homeland security (e.g., to detect thefts and assaults), healthcare (e.g., to monitor the actions of elderly patients and detect life-threatening symptoms), ecological monitoring (e.g., to analyze the behavior of animals in their natural habitats), and automatic sign-language recognition for assisting the speech-impaired. Although this problem has been tackled with limited success over the last decade or so, when storage and processing resources are severely limited and decisions need to be made in real time, no satisfactory system has been developed to date.

Another problem, which makes the task challenging, is action variability: the same action performed by different people may look quite different, e.g., people have different walking gaits. One source of difficulty is that different actions, such as walking and running, may sometimes look very similar due to the viewing-angle and the camera frame-rate. Give the video footage to a computer, however, and soon obstacles emerge.

Action Recognition Using Log-Covariance Matricesįunding: National Science Foundation (CISE-SNC-NOSS)īackground: A 5-year old child has no problem recognizing different types of human actions, such as, walking, running, jumping, waving, etc., from video footage captured by a surveillance camera.

Human activity detection matlab code