YOLO-DS: Application of One-Stage Instance Segmentation in the Dark at Construction Sites

Accurate instance segmentation can be used for scene reconstruction and analysis, so obtaining good image instance segmentation results is particularly important for construction scene safety management. However, there are still many challenges, such as challenging lighting conditions, diverse shapes of heavy equipment, etc. This paper proposes a novel YOLOv8-based network that innovatively addresses the instance segmentation task in the dark at construction sites, named YOLO-DS(Dark Segment). First of all, in order to solve the problem of low-light conditions, we introduced a low-light image enhancement network(LIEN) to adjust the lighting conditions of the images adaptively. LIEN can be integrated into the Yolov8 model for end-to-end training, avoiding the need for separate training of image enhancement and inconsistent goals of instance segmentation algorithms leading to the problem of good image enhancement but poor instance segmentation effect. Secondly, we have improved the training strategy of the original Yolov8 model to enable our model to more specifically solve the instance segmentation task, construction scene equipment deformation, and unclear target edges under dark light conditions. In summary, YOLO-DS shows extraordinary performance, achieving an average accuracy (mAP) of 55.5% on the public construction site data set MOCS, which is 2.8% higher than YOLOv8 and 1.5% higher than the SOTA(state-of-the-art) result on MOCS. At the same time, we achieved the mAP of 56.5% on the extracted lowlight dataset DarkMOCS, which is 4.5% higher than YOLOv8, fully verifying the effectiveness of our model on low-light images.