Lane-Line Segmentation Using YOLOv7

Lane-line segmentation has become increasingly significant in the contemporary times due to its compelling real-life applications in the diverse industries. The purpose of lane detection algorithms is to meticulously recognize the boundaries of the drivable-regions and carry out the localization. They form an enormously essential component of the Advanced Driver Assistance Systems (ADAS). Such systems would also help in significantly bringing down the accidents caused due to error by the human. In order to make such systems feasible, it's important to ensure they are reliable both in terms of inference speed and accuracy. Hence, this work proposes to use YOLOv7 model for the effective and efficient segmentation of the lane-lines from road-scene images. It follows the concept of trainable bag of freebies that allows the model to have both its performance as well as speed increase without negatively impacting each other. The Extended Efficient Layer Aggregation Network in its backbone network helps in improving the operation range while learning features. The performance of the model was compared to its predecessor YOLOv5. Results show that YOLOv7 outperforms with its optimal mean average precision of 92.4 for segmentation.