A Single-Stage Real-Time Object Detection Method Integrated Swin-Transformer and YOLO for Construction Management

Computer vision has significantly improved the management efficiency of safety, progress, efficiency, and quality in the construction process due to its real-time, high efficiency, low cost and ability to continuously and automatically recognize various classes of elements at the construction site. However, it is still a great challenge to accurately detect multiple classes of objects and small objects on the site in real-time due to the chaotic and large scope of the on-site construction environment and the concentration and variety of construction elements. Therefore, this study proposes a single-stage object detection method integrating Swin-Transformer and YOLOv5 for real-time detection of multiple classes of objects at construction sites. In the experiments, the v5n model with the smallest size is used for training, and the mAP@50 under this method reaches 80.57%, which is 10.81% higher than the original YOLOv5n model. It achieves an inference speed of 2 milliseconds per image, which meets the real-time requirements.