Social distance monitoring of site workers for COVID-19 using context-guided data augmentation, deep learning, and homography transformation

Because of the COVID-19 pandemic, many industries have developed efforts to minimize COVID-19's spread. For example, the construction industry in Melbourne practices social distancing and downsizes the number of workers on the job site. The surveillance system integrated with deep learning models has been extensively utilized to enhance construction safety. However, such 2D-based approaches suffer from occlusions, and the workers may not be accurately detected under this circumstance. To this end, this paper proposes a novel context-guided data augmentation method to enhance deep learning models' performance under occlusions. The context-guided method can automatically augment images by adding occlusions to the objects. Using this way, deep learning models can learn the object's features in various occlusion scenarios. Later, this method is validated by a real-time social distancing violation detection system. Specifically, this system utilizes a modified YOLOv4 model to detect workers by bounding boxes. Then, the DeepSORT algorithm is used to track the worker trajectories. Finally, homography transformation is used to calculate the distance between workers in each frame. The system has revealed robust results using the data augmentation method, and promising results indicate that the system can well support worker health during COVID-19.