Road Marking Detection in Challenging Environments using Optimized YOLOv7

Road marking detection area avails research with the application of computer vision and machine learning in the identification and analysis of various types of road markings such as lane markers, crosswalks, and road signs. It is very important to enhance safety on the roads by making these markings visible to drivers, cyclists, and pedestrians. Therefore, this study was designed to further enhance the efficiency and performance of the road marking detection models by incorporating the advanced AdamW optimizer and proposing a novel architecture known as Multi-Stage Cross-Convolutional Bottleneck Network-MSCCBN further-efficient in describing complex patterns and shapes from road markings. The testing was performed on a newly developed benchmark dataset containing 2,887 high-resolution images divided into training and test sets, with a total of 11 distinct road marking categories. Applying the Ceymo dataset and YOLOv7 detection framework, the result of the proposed model was an mAP50 score of 0.889, which means very high accuracy in object detection. This model indeed had a high precision of 0.886 and recall of 0.83, hence effectively minimizing both false positives and false negatives.