Automated Vision-Based Activity Identification for Demolition Operations

Demolition projects involve various types of heavy equipment (e.g., excavators, dump trucks, loaders, etc.). As such, the success of demolition projects is significantly dependent on heavy equipment operations. Prior studies have investigated heavy equipment productivity within the context of construction operations (i.e., earthwork) by tracking machine productivity through traditional approaches (e.g., manually tracking the duration of heavy equipment activities, etc.), which is a time-consuming, labor-intensive, and error-prone job. To facilitate research on heavy equipment productivity, recent studies employ artificial intelligence-based methods to automatically identify heavy equipment activities and measure productivity in construction operations. However, unlike earthwork activities where most of the tasks are relatively simple and repetitive, demolition activities are more complex and dynamic (i.e., related to structural demolition and material separation). Due to the varied nature of demolition activities, applying existing approaches to identify demolition activities is questionable. This study presents an automatic vision-based activity identification model based on three-dimensional Convolutional Neural Networks (CNNs), which can extract spatial and temporal features simultaneously. The proposed approach can recognize three excavator activities related to material separation (i.e., grabbing, swinging, and dumping) used in demolition operations. To develop the model, small-scale excavators were used to simulate a real-world demolition operation (i.e., separating materials), while two cameras were used to record videos of such experiments. Recorded video datasets were manually labeled and used to train the proposed model. Compared to construction projects, demolition projects are not relatively common. Therefore, it would have taken a while to collect the data from real-world demolition sites for training and validating the activity identification model. Through small-scale demolition simulations, the feasibility of the vision-based activity identification model was validated, which will support its application for full-scale demolition productivity improvement (i.e., by reducing the time and labor required for manual tracking of heavy equipment activities, monitoring the productivity of demolition operations, and enabling the development of timely and effective demolition strategies and productivity improvement measures accordingly).