Affiliations: [a] School of Mechanical Engineering and Automation, Northeastern University, Shenyang, China
| [b] Dandong Power Plant, HUANENG Power Intl Inc, Dandong, China
Correspondence:
[*]
Corresponding author. Yongfu Wang, School of Mechanical Engineering and Automation, Northeastern University, No. 3-11 Wenhua Road, Shenyang 110819, China. E-mail: [email protected].
Abstract: To solve the problem that the existing semantic segmentation method is challenging to balance the two indicators of reasoning speed and accuracy under low-light conditions, a lightweight real-time semantic segmentation method that can run on embedded vehicle terminals under low-light conditions, namely, LCG-BiSeNetV2 is proposed. Firstly, the Zero-DCE++ method is introduced to preprocess the low-light video frames; Secondly, by improving the module of the BiSeNetV2 method and optimizing the downsampling convolution in the encoder structure, the semantic feature extraction ability of the BiSeNetV2 method has been significantly improved; Finally, the self-made road scene semantic segmentation dataset and the CAMVID dataset were used for training and testing, the mIoU coefficient reaches 72.37% and 76.73%, and the reasoning speed reaches 148.91 Frames Per Second (FPS) and 136.75 FPS, which shows that the LCG-BiSeNetV2 method has achieved fast reasoning and precise semantic segmentation.
