Depth-Based Human Detection Considering Postural Diversity and Depth Missing in Office Environment

To realize robust human detection in an actual office work scenario, this paper proposes two ideas using top-view depth cameras. To deal with the changing geometric human shapes caused by body posture (e.g., sitting, standing, and crouching), we propose two features to describe the human upper-back shape, i.e., roundness and size of a height-continuous region. For alleviating the influences of partial loss of depth information caused by occlusions and by the absorption of infrared light, we propose an adaptive feature adjustment algorithm, which utilizes implicitly included information in the missing region. We implemented the proposed algorithm on a system with 13 depth cameras. Application to 100-hours (10 workdays) of actual office data demonstrated that the upper-back features complement the existing head-shoulder features. It also demonstrated that both of the proposals contributed to a more robust human detection and attained 97.7 % accuracy.