Scanning 3D Full Human Bodies Using Kinects

Abstract

Depth camera such as Microsoft Kinect, is much cheaper than conventional 3D scanning devices, and thus it can be acquired for everyday users easily. However, the depth data captured by Kinect over a certain distance is of extreme low quality. In this paper, we present a novel scanning system for capturing 3D full human body models by using multiple Kinects. To avoid the interference phenomena, we use two Kinects to capture the upper part and lower part of a human body respectively without overlapping region. A third Kinect is used to capture the middle part of the human body from the opposite direction. We propose a practical approach for registering the various body parts of different views under non-rigid deformation. First, a rough mesh template is constructed and used to deform successive frames pairwisely. Second, global alignment is performed to distribute errors in the deformation space, which can solve the loop closure problem efficiently. Misalignment caused by complex occlusion can also be handled reasonably by our global alignment algorithm. The experimental results have shown the efficiency and applicability of our system. Our system obtains impressive results in a few minutes with low price devices, thus is practically useful for generating personalized avatars for everyday users. Our system has been used for 3D human animation and virtual try on, and can further facilitate a range of home–oriented virtual reality (VR) applications.

Evaluation of Handheld Scanners for Automotive Applications

“…For some users, the cost also plays a significant role when it comes to selecting a particular device for the given application. For example, Tong J., et al employed a low cost 3D data-acquisition technique based on the Microsoft Kinect sensor [10]. This scanning system was used for capturing the complete 3D human body using multiple Kinect sensors.…”

Full 3D Reconstruction of Non-Rigidly Deforming Objects

“…After the advent of commodity RGB-D or 3D cameras based enhanced 3D reconstruction techniques for rigid objects [11,19,46,50], researchers have moved towards handling non-rigid deformations by proposing to construct complete and enhanced 3D models of mainly human subjects by fusing information from multiple views. This requires handling quasi-rigid motions between different views for which a global non-rigid registration is performed [38,57], or a model-to-part registration based on deformation graph [53] or Shape Completion and Animation of People (SCAPE) model [5] is used to avoid error accumulation [63,67]. The works of Cui et al [17] and Shapiro et al [24] are interesting in this regard as they try to tackle the limited-resolution of the data acquired from commodity 3D cameras as well.…”

Easy and Fast Reconstruction of a 3D Avatar with an RGB-D Sensor

“…In previous works, Weiss et al adopted a single Kinect to capture humans wearing undergarments in various poses and computed body shapes by fitting them into a SCAPE model [2,4]. Tong et al used three calibrated Kinects and a turntable to capture different body parts in the front and back sides and then reconstructed a 3D body shape [5]. Cui et al reconstructed the human body with a system where the subject turned continuously around 360 degrees before a static Kinect while maintaining an approximate “T” pose [6].…”

“…Based on the RGB-D data from Kinect, there has been a great deal of research on human reconstruction to achieve cost savings and greater efficiency [4,5,6,7,8,9], however, the final reconstructed human model is still far away to be acceptable. The existing research is primarily concerned with the following: (1) the use of fewer Kinects to increase modeling convenience and save on costs, as in [4,6,7,9]; (2) the use of fewer shape priors to reduce reconstruction complexity and obtain more of the actual body silhouette, such as in [5,8]; (3) the development of more sophisticated alignment algorithms to improve modeling accuracy, as in [5,7]; (4) the better quality RGB-D data inputs to obtain more model detail based on artificial denoising or a more reasonable capture strategy, as in [7,8,10,11,12]. Although previous works achieved a great deal in these aspects, it is still necessary to improve modeling quality and efficiency because of increasing demands on practicality.…”

“…Thus, there are a total of 18 frames that cover the full human body, which dramatically reduces the amount of captured frames compared to previous work. For instance, in the system by Cui et al [6], the user is turned around every 20 to 30 s to capture 10 frames within 0.5 s intervals; in the work of Tong et al [5], each Kinect acquired images at 15 frames per second about every 30 s; in the system of Li et al [7], the user was scanned from eight to nine views, and in each view, roughly 150 raw frames were captured. The greatly reduced amount of captured frames in our work helps reduce the computation loads for raw data analysis and thus speeds up the modeling procedure.…”