Lane Line Detection and Classification Based on Deep Learning

In the 21st century, with the progress of computer computing capability and the rapid development of machine learning, automatic driving technology is becoming more and more perfect. Nowadays, the field of autonomous driving technology has become a “strategic highland” for various vehicle companies. Lane detection is a key task in this field, which plays a vital role in the decision-making of automatic driving. At present, deep learning has made great achievements in various fields of computer vision, and it is also widely used in the field of lane detection. Compared with traditional image processing methods, the deep learning method is less affected by environmental and climate changes and has higher generalization ability, so the lane line recognition results output by this method are more stable. However, for most algorithms, although they have high accuracy, they cannot meet the high real-time requirements of automatic driving due to the heavyweight model and need to do a large amount of calculation. This paper presents a lane line detection-classification model: 1) An improved self-attention mechanism is added to ERFNet, the lightweight semantic segmentation model, to introduce higher-level semantic information to improve the prediction accuracy. 2) A lane line classification branch is added to the model. The original image and the output result of the semantic segmentation network are merged, and the new image obtained is used as the input of the classification branch. A convolutional neuron network is designed so that the model can predict the type of each lane line (solid line, dashed line). Our proposed model realizes the end-to-end lane line prediction and classification. To build this system, the CULane dataset has been labeled using 3 classes. The proposed method is validated on the CULane benchmark and reaches 73.1 F1-Measure.