Vision-Based Pose Forecasting of Construction Equipment for Monitoring Construction Site Safety

Abstract

Construction sites suffer from higher hazard rates compared to other occupational workplaces, which can be attributed to dynamic activities of construction equipment. Hence, it is important to track locations, poses, and movements of construction equipment for on-site safety monitoring. With the wide installations of surveillance cameras, previous studies focused on tracking locations of construction equipment from videos using computer vision techniques. However, the location of many construction equipment remains unchanged during operation while the equipment poses are varying constantly. The variation of equipment poses can cause severe hazards, such as the collision with surrounding workers and other equipment. To avoid such potential hazards, it is important to monitor and forecast the poses of equipment. So far, there is limited research on automatically forecasting poses of construction equipment, which is necessary to provide hazard alerts and prevent spatial conflicts. Therefore, a vision-based pipeline is proposed in this paper to automatically forecast dynamic poses of construction equipment based on historical on-site surveillance videos. The proposed pipeline firstly utilizes a deep learning -based equipment pose estimation model to estimate poses of construction equipment so as to obtain historical poses of construction equipment. Then, one type of recurrent neural network (RNN), namely Gated Recurrent Unit (GRU), is adopted to learn the temporal features of the generated sequential poses and forecast potential poses of construction equipment. To validate the proposed method, a dataset containing equipment keypoint-based poses is created and annotated for training the model. The experiment results based on our created dataset demonstrate the capability of the proposed pipeline.