Optical Flow Optimization Using Parallel Genetic Algorithm

Abstract

A new approach to optimize the parameters of a gradient-based optical flow model using a parallel genetic algorithm (GA) is proposed. The main characteristics of the optical flow algorithm are its bio-inspiration and robustness against contrast, static patterns and noise, besides working consistently with several optical illusions where other algorithms fail. This model depends on many parameters which conform the number of channels, the orientations required, the length and shape of the kernel functions used in the convolution stage, among many more. The GA is used to find a set of parameters which improve the accuracy of the optical flow on inputs where the ground-truth data is available. This set of parameters helps to understand which of them are better suited for each type of inputs and can be used to estimate the parameters of the optical flow algorithm when used with videos that share similar characteristics. The proposed implementation takes into account the embarrassingly parallel nature of the GA and uses the OpenMP Application Programming Interface (API) to speedup the process of estimating an optimal set of parameters. The information obtained in this work can be used to dynamically reconfigure systems, with potential applications in robotics, medical imaging and tracking.

Cite

@inproceedings{zavalaromero2011optical,
  title = {Optical Flow Optimization Using Parallel Genetic Algorithm},
  author = {Zavala-Romero, Olmo and Botella, Guillermo and Meyer-Baese, Anke and Meyer-Baese, Uwe},
  year = {2011},
  doi = {10.1117/12.883680},
  url = {https://doi.org/10.1117/12.883680},
}