Improved YOLOv8‐Based Bolt Loosening Detection for Tower Crane on Construction Sites

Tower cranes are indispensable heavy machinery on construction sites, and their structural integrity depends on numerous bolt connections. Bolt loosening poses severe safety risks such as potential collapse of a tower crane and casualties. Thus, early detection is critical. However, the traditional manual inspection is labor‐intensive, inefficient, and dependent on inspector experience. Therefore, the detection methods based on computer vision attract a lot of attention, but the detection result may be affected by lighting, obstacles, motion blur, and low resolution. Additionally, bolts typically occupy only 3%–5% of the image area, which presents a significant small‐object detection challenge. This work proposes an improved YOLOv8s‐based approach for robust bolt loosening detection in real construction environments. First, we incorporate a dilated reparameterization block (DRB) into the C2f module of the backbone network to enhance small‐object detection capability. Second, we replace the standard convolution module in the neck section with a grouped spatial convolution (GSConv) module to improve robustness against image blur. Finally, we integrate the VoV‐GSCSP (vector of vector‐grouped spatial convolutional spatial pyramid) module into the C2f module of the neck section to enhance multiscale feature fusion under occlusion. To validate the proposed method, we established a dataset of 1076 images captured from real construction site environments in Xi’an, China, including various challenging conditions such as different lighting (daylight, overcast, and artificial), viewing angles (frontal, side, and top‐down), and partial occlusions. The dataset is annotated with bounding boxes for two classes, i.e., tight_state and loose_state. Experimental results demonstrate that our improved model achieves 90.4% mAP@0.5, 88.7% precision, and 93.0% recall, outperforming the baseline YOLOv8s by 2.0% in mAP@0.5 (from 88.4% to 90.4%). Furthermore, the model achieves 435 frames per second (FPS) on GPU with a model size of 18.2 MB, which satisfies real‐time detection requirements and makes it suitable for application on edge devices or unmanned aerial vehicle (UAV) platforms.