Accurate Detection of the Workers and Machinery in Construction Sites Considering the Occlusions.

Traditional supervision of safety in construction sites is usually carried out manually, which may be inefficient and lacks real-time capability. In recent years, scholars have conducted the researches that utilize computer vision methods to perceive workers and machinery information to enhance safety supervision. However, due to complex backgrounds and occlusion issues in construction sites, the detectors still prone to generate missed detections. In order to automatically and accurately identify workers and machinery for accident prevention in construction sites, this paper proposes a novel object detection algorithm. In this method, Mosaic and MixUp data augmentation techniques are employed to obtain more training samples. An improved loss function is adopted to be the detection head to enable more accurate detection of the occluded objects. Experimental results show that the proposed algorithm exhibits superior performance in terms of accuracy and inference speed compared to the baseline detectors. It particularly performs better in detecting occluded objects, with an average precision improvement of 2.41%.