Detection of Limit Situation in Segmentation Network via CNN

The ability to detect limit situation is an essential element for ensuring safety in semantic segmentation task in self-driving system. In this paper, we study the detection of limit situations on the results of the image semantic segmentation network, and propose a framework consisting of convolution layers and fully connected layers. The mIoU value is deduced to evaluate a performance of semantic segmentation on the image obtained from the front vertical camera of the actual vehicle. The proposed network shows 90.51% accuracy in Hyundai Motor Group road image dataset for reasoning as a result of verification of the test set.