Dual Image Cropping Algorithm for Enabling Redundant ASIL-D Safe Lane Detection

The advancement of autonomous vehicles (AVs) to SAE level 4 has increased the focus on safety within autonomous driving systems. While AVs typically use diverse localization mechanisms for safe maneuvering, the vision-based localization on a lane is one of the most widely used. However, those lane detection systems often depend on a single camera, leading to safety concerns arising from potential random hardware faults. This paper provides an early indication to reduce this risk by proposing a novel cropping algorithm which detects the images overlap between two independent cameras. Both cropped images provide the input for two independent instances of lane detection algorithm, which are running in distinct hardware channels. The outputs of both lane detection instances are cross shared between distributed voters. Those voters can detect an eventual occurring random hardware fault in one of the channels. By doing so, the primary objective is to elevate the safety level of the system to the ISO26262 ASIL-D. Therefore, a quantitative analysis has been executed which compares the detected distance of the AV to the left and right lane marking of each lane detection algorithm. The system was able to produce promising results with a mean absolute error of just 3 cm, thereby indicating that it could be used to achieve the highest ASIL safety goal.