Fast Lane Detection Based on Attention Mechanism

Abstract

As an important subtask in the field of unmanned driving, lane detection has transitioned in recent years from traditional image processing to a deep learning-based neural network approach. However, since the early deep learning methods are based on semantic segmentation at the pixel level, their large network structures cannot meet the real-time requirements. To solve the real-time problem, a new network structure based on predefined rows, represented by UFAST, is proposed. In the architecture of this network model, the network parameters are significantly reduced, allowing the system's real-time performance to be satisfied. To improve the performance of recognizing lanes in the framework of this model, we introduce attention mechanism into the model by considering the habits of real human driving. Finally, we not only improve the performance of the model framework in non-ideal conditions such as poor lighting and vehicle occlusion by nearly 1.9%, but also increase the number of model parameters by less than 0.2% of the UFAST. We address our codes at https://github.com/APPZ99/Fast-Lane-Detection-Based-on-Attention-Mechanismon