Joint Detection And Activity Recognition Of Construction Workers Using Convolutional Neural Networks

Ghazaleh Torabi, Amin Hammad, Nizar Bouguila
Concordia University, Canada

DOI: 10.35490/EC3.2021.197
Abstract: Manually gathering information about activities on construction sites for project management purposes is labor-intensive and time-consuming. As a result, several works leveraged the already installed surveillance cameras to automate this process. However, the recent learning-based methods discretize continuous activities by assigning a single label to multiple consecutive frames. They do not fully leverage the contextual cues in the scene, and are not optimized end-to-end. A variation of the YOWO network, called YOWO53, is proposed in this paper to address these limitations. YOWO53 shows better classification and detection results over YOWO and allows using smaller input frames with real-time speed.
Keywords: computer-vision, construction productivity, convolutional neural networks, activity recognition
Pages: 212 - 219
Paper:
Contribution_197_final

Presentation Video: