Single Kinect Data Fusion for 3-D Modeling of Human Body

Three-dimensional modeling of the human body has become one of the important research topics in computer graphics. This is due to the importance of virtual representation of the human body in applications such as animation, computer games, virtual fitting room and cases etc. This has been obtained in the context of software and hardware developments in computer graphics. In this regard, three-dimensional modeling of the human body with low cost, high quality and accessible to everyone without the complexity and the need for specific expertise for processing is of great importance.
The aim of this paper is proposing a method for solving 3D human body modeling. Since the introduction of Kinect by Microsoft with features including low cost, no complexity, depth and color images production with a high frame rate and possibility of using in different lighting conditions, it could be a useful tool for our this purpose. But using the Kinect sensor for human body modeling confronts challenges such as raw data with low resolution and high noise, users movement during the scan, hidden areas and also a lack of accurate connection between depth and color data. In this regard, the idea of using the Kinect rotation motor in vertical angles in order to reduce the distance from the user to increase the quality of primary data was presented. The proposed non-rigid registration method was utilized for solving the problem of user instability during the scan. Also the sensor geometry calibration for accurate alignment of color and depth data was used.
In this paper, the procedure for 3D human body reconstruction is as follow: at the first, person is stayed on a specified distance from the Kinect and is scanned in the eight stations at three vertical angles. Then, colored point cloud are achieved by aligning color and depth images and extracting user data from background. Then rigid registration between sequential data stations is performed automatically. In order to solve the problem of instability during the scans, non-rigid registration is done between data station pairs. Finally, a general mesh was generated from the final point cloud and texture mapping is done to produce a realistic 3D body model.
The experimental results show that our rapid 3D human body modeling system has a high capability comparing to other similar systems. This system has a lower cost (less than 150 dollars), capacity of scanning in near distance without additional equipment such as rotation tables, and a higher quality of the final 3D model so that the details such as wrinkles and hair style is recognizable. The final 3D model generated from point cloud with about 4 mm density and 4mm noise thickness. Also the problem of low-quality in modeling of legs and shoes caused by a high movement during the scan, have been largely resolved. Therefore, we can generally say that the proposed method resolves the similar system’s weaknesses in data collection and processing steps. This makes our system proper for diverse applications and different environment.