Genetic optimized BP network method for camera calibration in binocular vision

Xingchen Hu, Honglei An, Hongxu Ma, Haibin Xie, Hongtao Xue2012 · DOI

Abstract

Camera calibration is the first step of positioning using binocular vision. Owning to the approximation capability of the neural network, a complex mathematical model needed by traditional calibration methods can be avoided. However the general neural network methods have their drawbacks to reduce its accuracy. This paper presents searching algorithm for the best structure and parameters of a neural network using an improved genetic algorithm (GA). The experiments show that this method can be used to establish a mapping between 2D coordinates and 3D coordinates directly and accurately, which is better than traditional calibration and general BP network methods.

Introduction

In binocular vision systems, such as those used in unmanned aerial vehicles (UAVs), camera calibration is essential for mapping 2D image coordinates into 3D real-world coordinates. Traditional methods rely on complex geometric models to account for lens distortion and alignment, often resulting in slow, non-linear computations. While Artificial Neural Networks (ANNs) offer a model-free alternative, they frequently suffer from two major issues: the lack of a theoretical basis for architecture selection and the tendency for standard training algorithms like Backpropagation (BP) to get trapped in local minima.

Methodology

The proposed GABP (Genetic Algorithm-Backpropagation) method addresses these limitations by optimizing the neural network in two distinct stages before final training:

  1. Topology Optimization: A GA with binary coding searches for the optimal network structure, determining the ideal number of layers and nodes.
  2. Weight and Bias Optimization: A second GA using real-number coding identifies optimal initial weights and biases. This "rough search" provides a superior starting point for the BP algorithm, preventing it from getting stuck in local minima.
Flowchart of the GABP optimization process, showing the dual stages of topology and connection weight optimization.

The final network takes a 4-element vector $(x_1, y_1, x_2, y_2)^T$ from the stereo cameras and outputs the 3D world coordinate $(X, Y, Z)^T$. Through this automated design, the researchers identified a 4-4-18-3 topology as the most effective architecture.

Experimental Setup and Results

Physical experiments were conducted using two GC660C digital cameras and a chessboard pattern moved to various heights. The GABP method was compared against standard BP networks and traditional geometric calibration methods.

The GABP method demonstrated significant improvements in both accuracy and training efficiency. It reached a Mean Squared Error (MSE) of 0.009 in just 446 epochs, compared to the standard BP's 0.670 MSE after 1280 epochs. Furthermore, the positioning time was reduced to 0.017 ms, a massive gain over the 5 ms required by traditional methods.

Comparison of error (CM) between standard BP and GABP networks across 100 sample points, highlighting the significantly lower and more stable error of the GABP approach.
Metric GABP (Proposed) Standard BP Traditional Method
Average Error 0.994 cm 3.677 cm 1.925 cm
Error Variance 6.659 25.25 17.82
Positioning Time 0.017 ms 0.017 ms 5.0 ms

Conclusion

The integration of Genetic Algorithms for structural and parameter optimization significantly enhances the performance of BP neural networks for camera calibration. The GABP method is approximately 3.7 times more accurate than standard BP and twice as accurate as traditional geometric methods, while offering real-time processing speeds suitable for high-speed UAV applications. Future work will focus on expanding the diversity of training samples to improve depth (Z-axis) accuracy and extending the method to dynamic tracking scenarios.

Summary updated

Abstract

Camera calibration is the first step of positioning using binocular vision. Owning to the approximation capability of the neural network, a complex mathematical model needed by traditional calibration methods can be avoided. However the general neural network methods have their drawbacks to reduce its accuracy. This paper presents searching algorithm for the best structure and parameters of a neural network using an improved genetic algorithm (GA). The experiments show that this method can be used to establish a mapping between 2D coordinates and 3D coordinates directly and accurately, which is better than traditional calibration and general BP network methods.
BibTeX
@misc{hu2012genetic,
  author = {Xingchen Hu and Honglei An and Hongxu Ma and Haibin Xie and Hongtao Xue},
  title = {{Genetic optimized BP network method for camera calibration in binocular vision}},
  year = {2012},
  doi = {10.2991/iccia.2012.261},
}
Figure