This thesis presents a deep-learning-based closed-circuit television (CCTV) sensor system design for anomaly identification in mass crowds. This system is really needed when there are large-scale demonstrations, making it easier for security forces to act early to prevent clashes. This application uses deep learning with the 8th version of the YOLO method to train data training. The data used is footage from CCTV recordings to be processed into reference data in order to detect anomalies. By using more than 2500 clips to be used as data training, it is hoped that it can become a reference for detecting clashes at demonstrations, then the data was trained for 50 iterations to produce accurate reference data for detecting anomalies in the demonstration. This application can be integrated with CCTV using the Real Time Stream Protocol (RTSP) method which can immediately take CCTV images for identification. This application is able to detect any anomalies in the occurrence of clashes. In conclusion, the deep learning-based CCTV sensor system design for anomaly identification in mass crowds offers a promising solution for intelligent video surveillance in public places. It is hoped that this application can help security forces in monitoring demonstration activities to avoid clashes and can monitor suspicious things.