Different approaches to optimize high-definition matrix headlights to improve computer vision

The contribution at hand presents and compares different online optimization approaches of dynamic illumination of matrix headlamps to improve automatic object recognition by neural networks. The approaches optimize, on the one hand, the network confidence and, on the other hand, the brightness of the image, the Weber contrast, and the gradient distribution on the image depending on the headlight beam pattern. The evaluation shows no objectively seen best cost function for the scenario studied, and selecting a cost function is a subjective decision. Optimizing the beam pattern to increase the confidence and intersection over union leads to inhomogeneous and subjectively disturbing beam patterns, and using contrast and gradient leads to similar results.