TY - JOUR
T1 - Camera marker networks for articulated machine pose estimation
AU - Feng, Chen
AU - Kamat, Vineet R.
AU - Cai, Hubo
N1 - Funding Information:
This research was funded by the United States National Science Foundation (NSF) via Grants CMMI-1160937 , CMMI-1265733 , CMMI-1265895 , and IIP-1343124 . The authors gratefully acknowledge NSF's support. The authors also thank Walbridge Construction Company, Eagle Excavation Company, and the University of Michigan Architecture, Engineering, and Construction (AEC) division for their support in providing access to construction equipment and job sites for experimentation and validation. The authors also thank undergraduate researchers Gabriel Bartosiewicz, Bradley Hecht, Jack Kosaian, Tracey Lo, Ritika Mehta, Andrea Mercier, Joshua Rios, and Matthew Stone for their assistance in conducting this research. Any opinions, findings, conclusions, and recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the NSF, Walbridge, Eagle Excavation, Purdue University, or the University of Michigan.
Publisher Copyright:
© 2018 Elsevier B.V.
PY - 2018/12
Y1 - 2018/12
N2 - The pose of an articulated machine includes the position and orientation of not only the machine base (e.g., tracks or wheels), but also its major articulated components (e.g., stick and bucket). To automatically estimate this pose is a crucial component of technical innovations aimed at improving both safety and productivity in many construction tasks. Based on computer vision, an automatic observation and analysis platform using a network of cameras and markers is designed to enable such a capability for articulated machines. To model such a complex system, a theoretical framework termed camera marker network is proposed. A graph abstraction of such a network is developed to both systematically manage observations and constraints, and efficiently find the optimal solution. An uncertainty analysis without time-consuming simulation enables optimization of network configurations to reduce estimation uncertainty, leading to several empirical rules for better camera calibration and pose estimation. Through extensive uncertainty analyses and field experiments, this approach is shown to achieve centimeter level bucket depth tracking accuracy from as far as 15 m away with only two ordinary cameras (1.1 megapixels each) and a few markers, providing a flexible and cost-efficient alternative to other commercial products that use infrastructure dependent sensors like GPS. A working prototype has been tested on several active construction sites confirming the method's effectiveness.
AB - The pose of an articulated machine includes the position and orientation of not only the machine base (e.g., tracks or wheels), but also its major articulated components (e.g., stick and bucket). To automatically estimate this pose is a crucial component of technical innovations aimed at improving both safety and productivity in many construction tasks. Based on computer vision, an automatic observation and analysis platform using a network of cameras and markers is designed to enable such a capability for articulated machines. To model such a complex system, a theoretical framework termed camera marker network is proposed. A graph abstraction of such a network is developed to both systematically manage observations and constraints, and efficiently find the optimal solution. An uncertainty analysis without time-consuming simulation enables optimization of network configurations to reduce estimation uncertainty, leading to several empirical rules for better camera calibration and pose estimation. Through extensive uncertainty analyses and field experiments, this approach is shown to achieve centimeter level bucket depth tracking accuracy from as far as 15 m away with only two ordinary cameras (1.1 megapixels each) and a few markers, providing a flexible and cost-efficient alternative to other commercial products that use infrastructure dependent sensors like GPS. A working prototype has been tested on several active construction sites confirming the method's effectiveness.
KW - Camera
KW - Error analysis
KW - Excavator pose estimation
KW - Marker
KW - Structure-from-motion
UR - http://www.scopus.com/inward/record.url?scp=85053477592&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85053477592&partnerID=8YFLogxK
U2 - 10.1016/j.autcon.2018.09.004
DO - 10.1016/j.autcon.2018.09.004
M3 - Article
AN - SCOPUS:85053477592
VL - 96
SP - 148
EP - 160
JO - Automation in Construction
JF - Automation in Construction
SN - 0926-5805
ER -