The Human Behavior Recognition Research Based On Kinect

Human behavior recognition based on compute vision was in a development stage, but it has an extensive prospect. For example, it can be utilized in security system and robot technology etc. Most of information provided from cameras are color images, and a majority of studies on human behavior utilized color images. With the further development of infrared technology, RGB-D cameras with new sensors entered the market. Depth images obtained from RGB-D camera were utilized in human behavior study, and human appearances were easier extracted from depth images than color images. The RGB-D camera named Kinect was used to collect data in this thesis. The main research works of this thesis can be described as follows:(1) Kinect was utilized to obtain bodies'data which extracted from skeleton images, and hand-held objects'data obtained from Kinect which extracted from depth images and color images. Hand-held objects were added into human behavior recognition. The algorithm showed a good performance on the behaviors which have the same pose but not the same hand-held object.(2) Human behavior recognition algorithm which had three-tier structure was proposed in this thesis. The bottom layer included the Kinect data streams. The middle layer was composed of key poses, body actions and hand-held objects. The top layer included states extracted from key poses and body actions, hand-held objects'labels. The states are subelements of behaviors. The three-tier structure can divided intricate behavior into simple parts. It decreased the difficulty of analysis. The common parts of behaviors can be extracted at the top layer analysis. The common parts utilized to model behaviors led to a better robust model of behaviors.(3) For the lab environment, this thesis realized a new approach to cut out efficient information. It is a effective way to extract key poses and body actions based on velocity of arms and legs' joints from the data stream. There is no need to analyze each frame and it reduce the amount of calculation.