Vision-Based Unmarked Road Detection with Semantic Segmentation using Mask R-CNN for Lane Departure Warning System

Road Lane detection plays an essential role in a Lane Departure Warning System (LDWS). LDWS application faces difficulty to detect lanes in complex environments such as unmarked roads. The vision-based lane detection for unmarked roads is difficult to perform due to vague and inconsistent visual features for object detection. In this study, the Mask Region Convolutional Neural Networks (Mask R-CNN) was reconstructed and trained with the collection of annotated unmarked road image datasets. The model was trained to enable the system to detect the unmarked road with various road conditions and perform semantic segmentation. About 2,000 samples of the unmarked road with various background conditions images were annotated using VGG image Annotator Tool. The trained model was evaluated to get the desired performance of unmarked roads segmentation for LDWS. The performance of deployed Mask R-CNN unmarked road detection model was 80.2% in accuracy using the Mean Average Precision method.