| description abstract | Site monitoring is indispensable for modern construction management. Contact approaches, represented by wearable devices, have problems such as privacy leaks and hindering working. Vision-based noncontact methods depend highly on light and environmental conditions, and have poor three-dimensional perception ability. To propose an all-weather noncontact activity identification approach on construction sites, four-dimensional (4D) millimeter-wave (MMW) radar is adopted in this study for the first time because of its excellent abilities of motion sensing, spatial sensing, and penetration. First, a feature processing method is proposed to convert the MMW signal to a seven-dimensional point cloud, which consists of the shape information (x, y, and z) and four attributes (Doppler′, SNR′, H, and V), representing the information of velocity, signal-to-noise ratio, height, and volume, respectively. Second, a novel deep learning framework is developed, which contains (1) one shape subnetwork, driven by the PointNet++ model, to capture the shape information of objects; (2) four attribute subnetworks to fully utilize the additional attribute features; and (3) a two-layer fusion module to combine all the outputs of the subnetworks. With precision of 0.963, recall of 0.961, and an F1 score of 0.962, the results show that the proposed method can accurately identify construction activities under different environmental conditions. It also can facilitate further development of MMW radar–based solutions for construction site analysis. | |
