Human detection with occlusion handling by over-segmentation and clustering on foreground regions
ACCV'12 Proceedings of the 11th international conference on Computer Vision - Volume 2
| Hi-index | 0.00 |
Previously, we put forth a new computer vision system for indoor well-being monitoring of elderly populations based on the use of multiple stereo camera pairs. That approach involves combining the strengths of image space with three dimensional volume element (voxel) space techniques. However, that system is fundamentally limited because it is based on color imagery from visible light cameras. In this article, we extend our prior research and consider a new, inexpensive infrared depth camera device, the Microsoft Kinect. Advantages, such as the ability to operate 24 -- 7 in low-to-no light conditions, and shortcomings are detailed. In addition, we discuss necessary algorithmic extensions to our mixed image and voxel space framework for the Kinect sensor. Experiments are performed in a laboratory designed to resemble an elders living quarter. Vision findings are evaluated using our prior high-level linguistic summarization of human activity work. Preliminary results indicate that the Kinect sensor does indeed work in a wider range of operating conditions and it can produce activity descriptions that match that of a human.