Proper use of personal protective equipment (PPE) is key to minimizing injuries from accidents in construction sites. Past research has attempted at identifying PPE from digital images and videos. However, fast detection of multiple PPE components (e.g., hard hat, vest), along with their colors in visual data has remained mostly underexplored. In this paper, the authors present a deep learning-based technique to detect PPE by finding similarities between an unknown image (a.k.a., query image) and a pool of known images (a.k.a., gallery images). In computer vision, finding similarities between images have many applications including search-by-example (i.e., finding images similar to the query image) and person re-identification (i.e., checking if a person in query image appeared previously in any of the gallery images). In this paper, a derivation of this technique, designed specifically for construction domain applications, is presented. Using this technique, a query image of a construction worker with unknown PPE is compared against a number of gallery images of construction workers with different combinations of known PPE (e.g., no hard hat or vest, only hard hat, only vest, and both hard hat and vest). Next, from the known PPE information (e.g., type and color) in the best-matched gallery images, the information on PPE in the query image is inferred. The paper illustrates a method that dynamically matches two images from top to bottom and finds the similarities based on the shortest distance between the local features. The proposed model is trained and tested on an in-house dataset containing 3,309 images of construction workers. Results show that the proposed method can identify PPE components with 90% accuracy, and colors of these components with 77% accuracy.
Publication Date: 2020
Source URL: Link to URL
Publication Types: Books, Reports, Papers, and Research Articles
Topics: Computer Vision; Construction Safety; Machine Learning; Mathematical Models; Personal Protective Equipment; Video Imaging Detectors; Worker Safety