Recursive Total Variation Filtering Based 3D Fusion

Abstract

3D reconstruction from mobile image sensors is crucial for many offline-inspection and online robotic application. While several techniques are known today to deliver high accuracy 3D models from images via offline-processing, 3D reconstruction in real-time remains a major goal still to achieve. This work focuses on incremental 3D modeling from error prone depth image data, since standard 3D fusion techniques are tailored on accurate depth data from active sensors such as the Kinect. Imprecise depth data is usually provided by stereo camera sensors or simultaneous localization and mapping (SLAM) techniques. This work proposes an incremental extension of the total variation (TV) filtering technique, which is shown to reduce the errors of the reconstructed 3D model by up to 77% compared to state of the art techniques.

A Versatile Decentralized 3D Volumetric Fusion for On-line Reconstruction

“…At present, various implementations of volumetric fusion [5,6,16,17] utilize modern CPU and GPU architecture to facilitate virtually boundless reconstruction with the help of hashed voxel grids. Therefore, in the presented system, we have extended a regularized implementation of [7] that provides controlled regularization to facilitate the implicit representation smoothing.…”

“…For instance, a mobile robot equipped with a depth sensor can scan the 3D geometry of a medium size room and store the output meshes in relatively good quality. Unfortunately, scaling the overall design to accommodate larger environments requires a complete 3D fusion-based reconstruction (such as InfiniTAM [5], FastFusion [6], and RFusion [7]) executed with higher computational and memory resource computing devices. Furthermore, their working principle limits active depth sensors to be bounded in an indoor environment, hence, scanning mechanisms should exhibit generic qualities to facilitate outdoor reconstruction.…”

Multi-Sensor Depth Fusion Framework for Real-Time 3D Reconstruction

“…Curless and Levoy [3] introduced a Signed Distance Function (SDF) based volumetric integration method which facilitates representation of sparse depth measurements in a dense bounded voxel space. Rajput et al [14] improved the underlying integration process by introducing a regularized variant least square based fusion which uses a semi-dense voxel space. This efficient utilization of voxel space reduced memory footprint as well as introduced smooth 3D surfaces.…”

“…This problem of reducing depth noise at the time of integration is addressed in [14] where a total variation filtering based regularization is employed to reduce the effects of noise in recursive manner. It is possible to treat both depth images (i.e.…”

“…where λ is the regularization parameter which controls the influence of neighbouring elements for SDF-signal. The mathematical solution to Equation 6 is used as-is from [14] where a recursive variant of the least squares system is derived and implemented. Finally, the estimated values of x are updated within the global voxel-grid.…”