Identifying and locating lane in the traffic environment is an important task for intelligent vehicle environment sensing system. Due to complex road environments, the lane detection task is still challenging. Because convolutional neural network (CNN) has the characteristic of local perception and lacks the ability to model the global context of the feature map, it's still hard for ordinary CNN to detect lane in complex traffic environment. In this study, we propose a fresh approach to detect lane base on an Efficient Attention Fusion Mechanism (EAFM). The main innovation of the model includes two aspects. Firstly, a Multi-scales Coarse Feature Extraction module (MCFE) is proposed to guide the feature learning of deep network by extracting rough feature maps at multiple scales; Secondly, we propose the EAFM to enhance the lane features and global information, so as to refine the result of lane detection. On Tusimple and CULane benchmark datasets, our model gets SOTA results.
EAFMNet: An Efficient Attention Fusion Mechanism for Lane Detection
2022-10-08
587240 byte
Conference paper
Electronic Resource
English
Wiley | 2023
|Vehicle lane changing prediction method based on attention mechanism
European Patent Office | 2022
|DOAJ | 2023
|