Real-time uncertainty estimation of stripe center extraction results using adaptive BP neural network

Abstract

Center extraction of stripe images is a key technique of line structured light sensors (LSLSs). As the original data for profile computation, the extraction results, which are highly affected by image quality, should be accurate and reliable. While, few works can be found for real-time uncertainty evaluation of the extraction results due to the diversity of center extraction methods and the complexity of uncertainty models. Here, a universal method is proposed to evaluate the uncertainty of center extraction results obtained from classical methods including, but not limited to, the gray gravity method (GGM), Gaussian fitting method (GFM) and Steger method. The proposed method is based on an adaptive BP neural network (ABPNN) where its weights and thresholds are adjusted automatically according to the width of each cross section profile. Experimental results show that the ABPNN can predict the uncertainty value of center extraction results accurately and efficiently.