Real-time reconstruction of depth sequences using signed distance functions

With the recent influx of inexpensive depth sensors such as the Microsoft Kinect, systems for 3D reconstruction and visual odometry utilizing depth information have garnered new interest. Often these processes are highly parallel and can be realized in real-time through parallel computing architectures for the GPU. We represent fused depth maps as a Truncated Signed Distance Function (TSDF) which is a grid of voxels that contain the distance to the nearest surface. Point clouds of subsequent captures are aligned to the TSDF volume through error minimization, taking advantage of the fact that surfaces are implicitly defined where the distance is zero. We present a new method for minimizing the error based on absolute orientation that does not require linearization or a weighting function. To evaluate the proposed method we compare the number of iterations required for convergence, time per iteration, and final alignment error with existing Gauss-Newton nonlinear minimization methods. While we use the Microsoft Kinect due to its fused depth and color capabilities, the alignment only requires depth and is applicable to current active fused lidar systems with VNIR, SWIR, MWIR or LWIR sensors.