Beam Recognition for Headlamps Aiming

In this paper, a novel approach is introduced to address the issue of vehicle headlamps aiming. Instead of handling it with traditional methods, it is considered as a beam pattern recognition problem. Unlike conventional techniques, which primarily rely on offline measurements of specific reference points, which are then projected onto a wall or panel, our method suggests an automatic aiming process using the Surround View Camera (SVC) via two main steps. First, checking the alignment of headlamps, where it detects whether it is “conform” or not according to the automotive lighting regulations. Second, estimating the deviation from optimal aiming regarding vertical and horizontal angles. A Deep Learning model is proposed to detect whether the headlamp is still within the regulation requirements, and another model for regression with a normalized loss function is designed to predict the errors of both orientations. Our two-step approach, tested with multiple architectures, shows high accuracy. To evaluate the performance of the proposed models, photometrical datasets of three different products are tested for both classification and regression. We computed several metrics on experimental results to evaluate both steps and obtained with the best considered architecture a mean accuracy of 94.46% for the classification part and a normalized error interval of 3.53% for the regression one.