Worker safety is still a serious problem, as accidents and deaths at work are unfortunately very common, especially in construction sites. To prevent such injuries, there is a great interest in video analysis solutions able to continuously monitor a construction site in order to verify the correct use of PPE by workers. In this paper we propose a method designed for verifying the correct use of helmet and protective vest, optimized to run directly on board of a smart camera, in order to use it also in temporary and mobile construction sites. The proposed approach solves many known problems in PPE verification systems, such as the detection at variable distances, the balancing of the number of samples available for the various classes of interest, a more accurate verification of the presence of the helmet and the management of challenging situations such as bald or hatted heads typically confused with helmets. To this aim, we conducted an ablation study showing the effectiveness of our design choices in terms of dataset preparation and classifier. The detection F-Score of 91.5 % in continuous monitoring, up to 94.0 % in more controlled access control scenarios, and the PPE recognition accuracy of 93.7 %, together with the capability to process 10 to 20 FPS on board of three different smart cameras, demonstrate the suitability of the proposed solution for real construction sites. An example video of the proposed system in action, integrated in a PPE verification product, is online available (https://www.youtube.com/watch?v=-fz25HYcFLo ).
Fast and Effective Detection of Personal Protective Equipment on Smart Cameras
Greco A.;Saldutti S.;Vento B.
2023-01-01
Abstract
Worker safety is still a serious problem, as accidents and deaths at work are unfortunately very common, especially in construction sites. To prevent such injuries, there is a great interest in video analysis solutions able to continuously monitor a construction site in order to verify the correct use of PPE by workers. In this paper we propose a method designed for verifying the correct use of helmet and protective vest, optimized to run directly on board of a smart camera, in order to use it also in temporary and mobile construction sites. The proposed approach solves many known problems in PPE verification systems, such as the detection at variable distances, the balancing of the number of samples available for the various classes of interest, a more accurate verification of the presence of the helmet and the management of challenging situations such as bald or hatted heads typically confused with helmets. To this aim, we conducted an ablation study showing the effectiveness of our design choices in terms of dataset preparation and classifier. The detection F-Score of 91.5 % in continuous monitoring, up to 94.0 % in more controlled access control scenarios, and the PPE recognition accuracy of 93.7 %, together with the capability to process 10 to 20 FPS on board of three different smart cameras, demonstrate the suitability of the proposed solution for real construction sites. An example video of the proposed system in action, integrated in a PPE verification product, is online available (https://www.youtube.com/watch?v=-fz25HYcFLo ).I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
