Video-based surveillance for safety in nuclear plants using deep learning

This work aims at developing a video-based system for surveillance purposes in nuclear plants. The system has to detect, track and correctly identify multiple people walking within the environment. Given the tracked trajectories, the doses received may be accounted for, by integrating the available dose rate map. This system is based in a former one that had been developed and achieved good results. This former system, however, operates offline. This new proposed system aims at replacing all the core methods implemented for detecting, tracking and identifying people, by using deep learning approach. Deep neural networks are suitable for these three tasks, and can replace all of them at once. The video data is provided by cameras installed in the nuclear plant room. Image samples obtained from these video data is used to train a deep neural network that is to be applied to unseen video data. Results are discussed. (author)