Construction Worker Posture Estimation Using OpenPose

Visual field observations are critical in construction project management. Cost reduction and increases in data usability have made RGB cameras more prevalent on construction sites, expanding capabilities of the modern construction manager. The eyes of a construction manager are irreplaceable tools which monitor and subjectively evaluate construction activities. The variety of "right" and "wrong" ways to complete a construction task make the automation of visual field operations a difficult task. Existing vision-based methods are limited to RGB-depth cameras which are rarely used on construction sites. This paper presents a method for using deep learning techniques to aid the automatic recognition of construction worker activities from RGB camera footage and to what extent can computer vision accurately detect construction worker activity. OpenPose human estimation algorithm was used to create 2D human pose samples of construction activities using RGB cameras. The samples are used to train and test a time distributed feed forward neural network with LSTM Keras model classifier. The results of this work in progress set the stage for improving the usefulness of visual data for project management in a variety of construction scenarios and other labor-intensive sectors. Monitoring the safety and productivity of construction workers will become more efficient allowing much-needed increases in our capacity to develop and maintain infrastructure.