Intelligent system for detecting faults in the industrial area

In this work, we propose a system capable of performing two tasks, recognition and prediction of human action using the surveillance cameras in the industrial environment to detect employees who violate the safety regulations by not wearing the safety cloths. This system is based on deep learning methods (convolutional neural networks, long-short term memory). For the human action recognition and prediction we proposed a new action representation, based on the human skeleton (body parts key-points), we used these points to create a vector containing the most important features of a descriptor (invariant to rotation, scale and position in the frame), and storing the spatial information of the video, we also used a position map to reduce its size and get very simple representation. Finally, we used LSTM network to preserve the temporal information, by training on these features using the dataset we created. The other part of this system is responsible of detecting employees who violate the safety regulation. So, in order to achieve this, we used the state of the art algorithm (You Only Look Once) for object detection and we adapted it to our dataset that contains four classes (Wearing-helmet, Without-helmet, Wearing-boot and Without-boot).