Calibrate Multiple Consumer RGB-D Cameras for Low-Cost and Efficient 3D Indoor Mapping

Abstract

Abstract:Traditional indoor laser scanning trolley/backpacks with multi-laser scanner, panorama cameras, and an inertial measurement unit (IMU) installed are a popular solution to the 3D indoor mapping problem. However, the cost of those mapping suits is quite expensive, and can hardly be replicated by consumer electronic components. The consumer RGB-Depth (RGB-D) camera (e.g., Kinect V2) is a low-cost option for gathering 3D point clouds. However, because of the narrow field of view (FOV), its collection efficiency and data coverages are lower than that of laser scanners. Additionally, the limited FOV leads to an increase of the scanning workload, data processing burden, and risk of visual odometry (VO)/simultaneous localization and mapping (SLAM) failure. To find an efficient and low-cost way to collect 3D point clouds data with auxiliary information (i.e., color) for indoor mapping, in this paper we present a prototype indoor mapping solution that is built upon the calibration of multiple RGB-D sensors to construct an array with large FOV. Three time-of-flight (ToF)-based Kinect V2 RGB-D cameras are mounted on a rig with different view directions in order to form a large field of view. The three RGB-D data streams are synchronized and gathered by the OpenKinect driver. The intrinsic calibration that involves the geometry and depth calibration of single RGB-D cameras are solved by homography-based method and ray correction followed by range biases correction based on pixel-wise spline line functions, respectively. The extrinsic calibration is achieved through a coarse-to-fine scheme that solves the initial exterior orientation parameters (EoPs) from sparse control markers and further refines the initial value by an iterative closest point (ICP) variant minimizing the distance between the RGB-D point clouds and the referenced laser point clouds. The effectiveness and accuracy of the proposed prototype and calibration method are evaluated by comparing the point clouds derived from the prototype with ground truth data collected by a terrestrial laser scanner (TLS). The overall analysis of the results shows that the proposed method achieves the seamless integration of multiple point clouds from three Kinect V2 cameras collected at 30 frames per second, resulting in low-cost, efficient, and high-coverage 3D color point cloud collection for indoor mapping applications.

Indoor Topological Localization Using a Visual Landmark Sequence

“…As mobile computing and smartphones are becoming readily accessible, there have been attempts to use smartphone cameras for indoor localization [17,[24][25][26][27][28][29][30][31][32][33][34][35][36]. These methods exploit computer vision techniques to estimate people's location and mainly fall into two categories: image retrieval-based methods and 3D model-based methods.…”

Audio Guide for Visually Impaired People Based on Combination of Stereo Vision and Musical Tones

“…In systems that use object identification, it is necessary to develop a classifier construction step and their mapping in the scenario [10]. On systems using distance indication, algorithms calculate the area where it has the most significant volume of data, indicating the presence of an object, whether this processes 2D (only perception of the ground level) or 3D images (ground level and vertical) [25]. The model adopted in this paper identifies the presence of obstacles by applying image processing without the use of classifiers, in two 2D matrices representing the horizontal and vertical planes.…”

Enhancement of RGB-D Image Alignment Using Fiducial Markers

“…The introduction of low-cost RGB-D cameras created an opportunity for 3D reconstruction of scenes to be performed at consumer-level. The increasing popularity of these sensors promoted the research of their use to reconstruct 3D indoor scenes [14][15][16]. Their manoeuvrability, permitting hand-held operation and allowing the user to get closer to parts of the scene and capturing them from different angles, proves to be an advantage in indoor environments with high probability of object occlusion, when compared with fixed high resolution scanners, which usually require more space and fixed poses to operate.…”