Event detection system for the penitentiary institutions using multimodal data and deep networks

The aim of the paper is to present the distributed system for the unwanted event detection regarding inmates in the closed penitentiary facilities. The system processes large number of data streams from IP cameras (up to 180) and performs the event detection using Deep Learning neural networks. Both audio and video streams are processed to produce the classification outcome. The application-specific data set has been prepared for training the neural models. For the particular event types 3DCNN and YOLO architectures have been used. The system was thoroughly tested both in the laboratory conditions and in the actual facility. Accuracy of the particular event detection is on the satisfactory level, though problems with the particular events have been reported and will be dealt with in the future.