| dc.description.abstract | In homes, the child accidents are occurred due to the limitation and block of the home construction and furniture as well as the children‘s own risk behaviors. In this paper, we propose a home environment monitoring system to survey children in homes and other indoor environments to avoid the child accidents. At first, we define three different regions in the house floor: active, inactive, and inhibition regions. Then the proposed system monitors the children’s location and behavior in time. Finally, the proposed system raises warning if there are abnormal or dangerous behaviors of children in various regions. The warnings are different according to different regions. In this study, the considered child’s behaviors include walking, running, falling/squatting, sitting, and standing on unsafe location.
In the proposed surveillance system, we use a 360-degree fisheye camera installing on the ceiling to capture 360-degree surrounding images. We use the codebook background subtraction method to extract the body silhouettes. In the multi-layer background model, we can handle moving background with periodic variation and illumination variations.
In falling detection, we use object directions based on the principal component analysis (PCA) and the height of body trapezoid bounding box as features to determine falling. We prefer to use trapezoid bounding box technique, because it is faster. However, in the center area of the omni-directional images, the trapezoid bounding box is unreliable; thus, we distinguish an image into two concentric regions: inner and outer concentric regions. In outer region, we use the height of trapezoid bounding box to decide falling. In inner region, we use PCA to find the first and second principal components and their ratio to determine falling. In order to increase the accuracy, we add a condition that a falling behavior must be followed by a period of a still status. In squatting detection, we use the same way as falling detection in outer concentric region. In inner concentric region, we can’t detect squatting by body shape, since the self-cover of body is serious. In running detection, we compute the foot position, and use consequent images to estimate the body moving speed. The moving speed is influenced by the position in images; thus, a pre-calibrated process on the acquired speed data is conducted.
In order to make this system be adequate for home environment, the whole home floor is distinguished into three areas: (i) active area (ex, floor), (ii) inactive area (ex, chair and sofa), (iii) inhibition area (ex, kitchen and balcony). Different behaviors only permit in different areas. The abnormal behaviors defined in the active area are running and falling. In the inactive area, the abnormal behavior is standing on chair or sofa. In the inhibition area, no any behavior is permitted; that is, once a foot position is detected in the area, the warning is then raised.
Many experiments were conducted; according to the experimental results, we find that the accuracy rate of falling detection is 86.7% with only using geometric characteristics of body appearances and 90.6% with adding still status determination. Children running detection rate is 95%. The detection rate of standing and sitting in inactive area is 100%. The detection rate of getting into inhibition area is 100%. The results reseal that the proposed system is effect and practical for the related applications.
| en_US |