Stereoscopic 3D Reconstruction using Motorized Zoom Lenses
within an Embedded System
Pengcheng Liu, Andrew Willis, Yunfeng Sui
UNC - Charlotte, 9201 University City Blvd., Charlotte, NC 28223
Stereoscopic reconstruction systems are found in a number of environments and have been under development for over 20 years. Ongoing changes in imaging technologies have driven continual theoretical development and technical variations on the stereoscopic reconstruction problem. This paper describes a novel stereoscopic 3D reconstruction system meant to act as an 3D sensing payload for a terrestrial robot.
Novel theoretical and technical aspects of the system are tied to two aspects of the system design that deviate from typical stereoscopic reconstruction systems: (1) incorporation of an 10x zoom lens (Rainbow-H10x8.5) and (2) implementation of the system on an embedded DSP/FPGA system.
Hardware: The system is implemented on a mixed DSP/FPGA system consisting of a Blackfin DSP (BF537-ezkit) and a Xilinx Spartan3 FPGA. The DSP is tasked with orchestrating data flow through the system and complex computational tasks. The FPGA acts as an interface between the DSP and the system devices which include the camera CMOS sensors and the servo motors which rotate (pan) each camera. The entire system runs on an embedded version of the Linux operating system called μClinux which has 64MB of available memory.
Software: Calibration of the camera pair is accomplished using a collection of stereo images that view a common chess board calibration pattern for a set of pre-defined zoom positions. Calibration settings for an arbitrary zoom setting is then obtained by interpolation of the camera parameters. Classical techniques are use to rectify images using the estimated calibration parameters. Dense stereo matching is performed on the stereo image pairs using a custom adaptation of a dynamic programming algorithm proposed by Filho & Aloimonos in 2006 which requires little memory and provides high performance compared to other techniques while sacrificing accuracy by limiting the number of paths considered in the disparity space. Subsequent 3D surface reconstruction is accomplished by classical triangulation of the matched points from the disparity map.
The paper includes descriptions of and results for our solutions to the following problems: (1) automatic com
putation of the focus and exposure settings for the lens and camera sensor, (2) calibration of the system for various zoom settings, (3) automatic control of the extrinsic parameters as a function of the zoom setting to ensure the camera pair has an overlapping field-of-view, (4) our adaptation of the dense matching algorithm and (5) stereo reconstruction results for several free form objects. Preliminary results for reconstructing a small figurine are provided in figure 1.
Figure 1. (a) shows an image of the embedded system. (b,c) show images of the figurine from the left (b) and right (c) cameras respectively. (d) is a preliminary 3D reconstruction of the figurine.