Accurate eye-like segmentation in a heavily untextured contrasted scene

Automatic pattern recognition is a hard task to carry out when shapes or textures have to be detected. This can be even more difficult when accurate photometric measures are required. Nevertheless, most of times the feasibility to have a measurable ground truth at our disposal (it maybe is achievable using other sensors) gives us a reference point that makes the researcher task easier. In this paper, we present a method to segment automatically a light-shadow line in a very high-contrasted scene, in the context of an industrial automotive application, where the ground truth is the line as being perceived by sight from an experienced operator. After segmenting the line, some measures have been achieved related to accuracy and precision of a reference parameter of the line (the "elbow"), using an industrial prototype integral with the headlamp to be tested. The experiments prove how the method we developed is able to detect perturbation of the headlamp beam in pitch and yaw lower than 1/10deg, this representing an excellent outcome.