Intelligent pixel-level pavement marking detection using 2D laser pavement images

Zishuo Dong,Hang Zhang,Allen A. Zhang,Yang Liu,Zhihao Lin,Anzheng He,Changfa Ai

DOI: https://doi.org/10.1016/j.measurement.2023.113269

IF: 5.6

2023-09-01

Measurement

Abstract:This paper proposes a robust semantic segmentation algorithm named as Marking-DNet to implement pixel-level recognition of pavement markings. The proposed Marking-DNet presents an improved encoder-decoder architecture based on DeepLabV3+. Different from DeepLabv3+, feature maps of four different scales are imported from the encoder into the decoder of Marking-DNet, resulting in more levels of information exchange. Additionally, the Object-Contextual Representation and the Convolutional Block Attention Module are both employed in Marking-DNet to conduct contextual learning more efficiently and implement spatial and channel attention explicitly. The F-measure and Intersection-Over-Union attained by the Marking-DNet on 1500 testing images are 93.59% and 87.96% respectively. Experimental results demonstrate that Marking-DNet outperforms six state-of-the-art semantic segmentation models in detecting pavement markings using both private and public image datasets. According to the performances on 5,506 pavement images, it is found that Marking-DNet can detect various complex pavement markings accurately, except for some heavily-worn markings.

engineering, multidisciplinary,instruments & instrumentation