Deep Feature Fusion Network for Lane Line Segmentation in Urban Traffic Scenes

As autonomous driving technology continues to advance at a rapid pace, the demand for precise and dependable lane detection systems has become increasingly critical. However, traditional methods often struggle with complex urban scenarios, such as crowded environments, diverse lighting conditions, unmarked lanes, curved lanes, and night-time driving. This paper presents a novel approach to lane line segmentation in urban traffic scenes with a Deep Feature Fusion Network (DFFN). The DFFN leverages the strengths of deep learning for feature extraction and fusion, aiming to enhance the accuracy and reliability of lane detection under diverse real-world conditions. To integrate multi-layer features, the DFFN employs both spatial and channel attention mechanisms in an appropriate manner. This strategy facilitates learning and predicting the relevance of each input feature during the fusion process. In addition, deformable convolution is employed in all up-sampling operations, enabling dynamic adjustment of the receptive field according to object scales and poses. The performance of DFFN is rigorously evaluated and compared with existing models, namely SCNN, ENet, and ENet-SAD, across different scenarios in the CULane dataset. Experimental results demonstrate the superior performance of DFFN across all conditions, highlighting its potential applicability in advanced driver assistance systems and autonomous driving applications.

@article{Nguyen2023,
  title     = {Deep Feature Fusion Network for Lane Line Segmentation in Urban Traffic Scenes},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {14},
  number    = {6},
  year      = {2023},
  publisher = {The Science and Information Organization},
  author    = {Hoanh Nguyen},
  doi       = {10.14569/IJACSA.2023.0140646},
  url       = {https://doi.org/10.14569/IJACSA.2023.0140646}
}