A Methodology for Extracting Standing Human Bodies from Video Stream

Abstract

Imaging of human body segments is demanding task which supports many applications such as understanding of scenes and recognition of activities. A bottom-up technology for extracting human bodies automatically from single image, in case of almost upright position, is the available technique in cluttered environments. The dimension, position and face color are used for localizing human body, model construction of upper and lower body as per anthropometric constraints and skin color calculation. A highest level pose can be extracted by combining different levels of segmentation granularity. Jointly estimating the foreground and background during the body part search phase gives rise to the segments of human body, that alleviates the need for shape matching exactly. We have now proposed a system in which human body is extracted from video and from complex pose scenes in motion vectors. In the future, we wish to deal with more Complex poses, without depending on strong pose prior. More masks can be incorporated to deal with problems such as missing of extreme regions like hair, gloves and shoes, in the search of these parts, but care should be taken in keeping the computational complexity from rising excessively. 40 images (43 persons) from dataset of INRIA persons and 163 images from the dataset of ‘lab1’ are used for measuring the performance of our algorithm, where the accuracies measured are 89.53% and 97.68% respectively. Qualitative and quantitative experimental results prove than our technique outperforms state-of-the-art interactive and hybrid top-down/bottom-up approaches.