Research On Chessboard Corner Detection Algo- Rithm Based On Statistical And Machine Learning

Automatic detection of chessboard corner is essential for camera calibration,and is widely used in surface reconstruction,visual tracking and pose estimation.Most chessboard corner detection algorithms need to manually select the suitable thresholds for fast false points suppression.Manually adjusting the thresholds will interfere with the automatic detection of chessboard corner,thus limiting the use of chessboard corner.In order to reduce the manual intervention in chessboard corner detection,to meet the real-time requirements of automatic corner detection,combining statistics and machine learning algorithms,this paper presents a cascade chessboard corner detection algorithm.The main work of this paper includes:(1)The spectral characteristics of the circular sampling sequence at the chessboard corner are analyzed,and a chessboard corner response function which characterizes the spectral characteristics of the circular sampling is defined,so that each pixel response value can represent the attributes of the chessboard corner.A method of rough selection of chessboard corner based on spatial distribution characteristics of corner response value is proposed.(2)Based on the Gaussian distribution hypothesis,the cumulative distribution function reflecting the chessboard corner response of image noise is derived.The method of automatically determining the threshold for eliminating noise points based on the distribution function is established,and an image noise estimation method based on superpixels is proposed.(3)The distribution characteristics of the intensity difference between the circular sampling center and the sampling sequence under different conditions are analyzed.Combined with the chessboard corner response value,a chessboard corner descriptor is proposed,and a fast stripe center point elimination method using the descriptor is established.(4)The correctness of the proposed theory and method is verified by chessboard corner detection experiments.The result shows that this method is robust against different image contrast,noise level and perspective distortions.This method can effectively remove most false points,yielding at least 99.2% recall and 97.8% precision for real-world images.