Research On Extraction Of Human Action In Video

Extraction of human action in video is an essential issue in computer vision area since it has many applications such as human visual video surveillance,human computer interaction and so on.The main task for human action extraction is to estimate human pose of each frame in video,and output the estimated results of each frame.It makes the pose estimation a tricky problem because of the complex image background,the diversity of the human action and cloths.This thesis mainly studied the three points: human detection in video,segmentation of foreground and background with Grab cut algorithm,the detection of human body parts.Main contributions of this thesis are follows:As the uncertain of human location in video,this thesis proposes the three-frame difference algorithm based on background and acquires the exact human location.Experiments shows that the new algorithm eliminates the ‘empty' effect of the original three-frame difference method,this thesis adjusts the size of the detection window adapt to preset parameters of Grab cut algorithm based on the proportion of human in the detection window.As the error segmentation of Grab cut algorithm due to the impact of complex background,this thesis proposes a new Grab cut segmentation based on the background,the background subtraction obtains binary images and replace the original ones by Grab cut segmentation.Experiments show that accurate binary images get exact segmentation results.The pose estimation of pictorial structure based on the color feature can cause the human body parts to be misidentified because of the complex background.Considering the texture of human body or skin is different from the background,this thesis put forward a new pictorial structure based on the color and texture features.Experiments show that when the body parts and the background own the similar color features,the improved method has higher accuracy.Using the results completed in this thesis to extract the human pose in the video.Firstly,detect the position of the human body in the video.Secondly,estimate human pose of each frame in the video.Finally,output the pose estimation results for each frame.Repeat the three processes mentioned above until finish the whole video.The improved algorithm improves the automation and accuracy of human pose estimation,and extract the human action effectively.