The system is able to recognize sequences where the

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: he system is able to recognize sequences where the gait of the subject in the input sequence differs considerably from the training sequences on which it has been modeled. Since the system uses the di erence in co-ordinates of the head as its feature vectors, it is able to recognize actions for people of varying physical stature, i.e., tall, short, thin, fat, etc. Hence the system can recognize the bending down action of both a short as well as tall person. For instance, in a sideways bending action where the head traces a curve whose radius is roughly equal to half the length of the body, the size of the radius itself may di er depending on the height of the person; however, the shape of the curve traced out in each case is the same. Thus, our system is not sensitive to the physical stature of the subject. Much work has been done in the area of human activity recognition. Cai and Aggarwal 1 discuss the di erent approaches used in the recognition of human activities. They classify the approaches towards human activity recognition into state-space and template matching techniques. Liao et al 2 discuss methodologies which use motion in the recognition of human activity. Ayers and Shah 3 have developed a system that makes context-based decisions about the actions of people in a room. These actions include entering a room, using a computer terminal, opening a cabi- net, picking up the phone, etc. Their system is able to recognize actions based on prior knowledge about the layout of the room. Davis, Intille and Bobick 9 have developed an algorithm that uses contextual information to simultaneously track multiple, non-rigid objects when erratic movements and object collisions are common. However, both of these algorithms require prior knowledge of the precise location of certain objects in the environment. In 3 , the system is limited to actions like sitting and standing. Also, it is only able to recognize a picking action by knowledge of where the object is and tracking it after the person has come within a certain distance of it. In 7 , Davis uses temporal plates for matching and recognition. The system computes history images MHI's of the persons in the scene. Davis 7 computes MHI's for 18 di erent images in 7 di erent orientations. These motion images are accumulated in time and form motion energy images MEI's. Moment-based features are extracted from MEI's and MHI's and employed for recognition using template matching. Although template matching procedures have a lower computational cost, they are usually more sensitive to the variance in the duration of the movement. A number of researchers have attempted the full three-dimensional reconstruction of the human form from image sequences, presuming that such information is necessary to understand the action taking place 10, 6, 14 . Others have proposed methods for recognizing action from the motion itself, as opposed to constructing a three-dimensional model of the person and then recognizing the action of the model 11, 4...
View Full Document

This note was uploaded on 04/12/2013 for the course ECON 2781631 taught by Professor Coop during the Spring '13 term at Emmanuel.

Ask a homework question - tutors are online