Fast lane detection for extracting spatial location information

Xiao Tan, Sheng Li, Hua Yan

Abstract: At present, even though great progress has been made in lane detection based on deep learning method in complex scenarios, there is still room for improvement in the real-time performance of most models. Row-wise classification method is the current mainstream method to improve the real-time performance of the model. It makes a trade-off between accuracy and speed. However, many models based on the row-wise classification method are not strong enough to extract spatial contextual information, This hinders the recognition of lanes. Inspired by Feature Pyramid Networks, we propose a simple and lightweight framework based on row-wise classification method: SIE-Net. The method can fully extract the spatial position information in the image. The framework can fuse the semantic information contained in the deep feature map and the spatial information contained in the shallow feature map. Then dilated convolution is used in the feature extraction process, which increases the receptive field of the model and extracts more global information in the image. Meanwhile, channel attention mechanism is used in the feature extraction process. It can give greater weight to the channel containing the structure information of the lanes. Finally, the experimental results demonstrate the effectiveness of the proposed method in terms of accuracy and speed on two popular Tusimple and CULane benchmark datasets.