Weitao Yang

Weitao Yang

Auto-generated profile, based on publicly available information

Weitao Yang

China Construction Seventh Engineering Bureau Co., Ltd., China

Research Focus & Metrics

Representative Publications

View All Publications
24 September 2023
Verified
Conference Paper
Open access
ZHZhenguo Hou
Zhenguo Hou
ZZZhongshan Zhang
Zhongshan Zhang
WYWeitao Yang
Weitao Yang
HLHao Luo
Hao Luo
ZWZheng Wang
Zheng Wang
CLChong Long
Chong Long
Scene monitoring and automatic interpretation in construction site scenarios play a crucial role. However, the changes in illumination and weather conditions of construction sites have a significant impact on the imaging quality of visual cameras. The effectiveness of existing semantic segmentation methods is extremely decreased without considering this problem. To solve this issue, we propose the multi-stage feature reconstructed semantic segmentation method for all-day monitor in construction sites. This method aims to improve the robustness against changes in the external environment. Specifically, the self-calibrated module (SCM) is adopted to reconstruct the input image and calculate the effective encoded features. SCM is embedded into the network for several stages to reconstruct multi-level features. Then, the higher-level features are aggregated and enhanced along the channel dimensions, which could obtain more context information. The higher-level is gradually combined with lower-level detailed features for decoding process. The images in construction sites are randomly augmented from different aspects, which could supervise the training process to achieve robust semantic segmentation with the supplement of image reconstruction. The qualitative and quantitative experiments show that compared with several state-of-the-art semantic segmentation methods, the proposed method could better adapt to changes in the environmental conditions and achieve all-day segmentation.