Smart construction scheduling monitoring using YOLOv3-based activity detection and classification

Increasing efficiency and adhering to a schedule are prominent issues faced by many construction projects. Identifying areas where productivity is low would automatically be a helpful tool for managers. This research aims to analyze and compare the efficiency and accuracy of different computer-vision based activity recognition algorithms that are used on construction sites. The authors then propose a method which involves the use of YOLOv3 to perform activity recognition on construction sites and compare the accuracy of our method to existing algorithms. The algorithms for comparison are selected on the basis that: (1) they incorporate various state-of-the-art activity recognition techniques, such as bounding-box predictions and skeleton-models; and (2) they are relatively recent implementations. The authors trained the model using a data-base consisting of 4 activities with frames from 20 videos for each. The dataset was created by extracting frames from the videos and labelling the activities that are taking place in each video. The authors then use the aforementioned activity classification method to propose a smart schedule monitoring system that automatically updates start and finish times of individual activity conducted in a construction project based on the activities that are detected. This computer-vision based approach to provide automatic and real-time updates to the construction schedule is expected to improve worker productivity and shorten construction project timelines.