Visual Design and Engineering Laboratory

Carnegie Mellon University

Recognizing Planar Kinematic Mechanisms from a Single Image Using Evolutionary Computation

Paper

[Paper PDF]    

[Publisher Version]    

[Data]    

Abstract

In this paper, a method is presented that automatically recognizes kinematic mechanisms from textbook images using an evolutionary algorithm to complement computer vision techniques for object detection. Specifically, a nondominated sorting genetic algorithm (NSGA-II) is used to optimize the number and position of mechanical joints in an image and corresponding joint connections (i.e. rigid bodies) such that Pareto front solutions maximize image consistency and mechanical feasibility. A well-known object detector is used as an example method for locating joints, and local image features between pairwise detected joints are used to predict likely connections. The performance of the algorithm using these specific vision techniques is compared to a parameterized detection scheme in order to decouple the efficacy of the object detector from the evolutionary algorithm. Experiments were performed to validate this approach on selected images from a custom dataset, and the results demonstrate reasonable success in both accuracy and speed.

Citation

Matthew Eicholtz, Levent Burak Kara, Jason Lohn. (2014). Recognizing Planar Kinematic Mechanisms from a Single Image Using Evolutionary Computation. In Proceedings of the 2014 Conference on Genetic and Evolutionary Computation (GECCO ’14), July 12-16, Vancouver, BC, Canada, ACM, New York, NY, pp. 1103-1110.

@inproceedings{eicholtz:2014:gecco,
 author = {Eicholtz, Matthew and Kara, Levent Burak and Lohn, Jason},
 title = {Recognizing Planar Kinematic Mechanisms from a Single Image Using Evolutionary Computation},
 booktitle = {Proceedings of the 2014 Conference on Genetic and Evolutionary Computation},
 series = {GECCO '14},
 year = {2014},
 isbn = {978-1-4503-2662-9},
 location = {Vancouver, BC, Canada},
 pages = {1103--1110},
 numpages = {8},
 url = {http://doi.acm.org/10.1145/2576768.2598354},
 doi = {10.1145/2576768.2598354},
 acmid = {2598354},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {computer vision, evolutionary multiobjective optimization, kinematic simulation, object recognition},
}