Driven by the rapid strides in information tech- nology, video surveillance systems have seamlessly integrated themselves as pivotal components within contemporary urban security and protection frameworks. This holds particularly true for locations like prisons, where surveillance cameras are ubiquitously deployed. However, as the surveillance network continually expands, these cameras bring not only convenience but also generate an extensive volume of monitoring data. This situation presents significant challenges related to data storage, analysis, and retrieval. Integrating intelligent video analytics technology into a smart monitoring system can effectively oversee and proactively alert for anomalous events or behaviors. This area represents a prominent avenue of research within the surveillance domain. In this study, deep learning techniques are employed, utilizing the cutting-edge instance segmentation frame- work known as Mask R-CNN. The methodology the authors adopt encompasses the training of a fine-tuning network using the dedicated datasets. This network showcases its adeptness in efficiently recognizing objects within video frames, all the while generating precise segmentation masks for each identified instance. Empirical findings underscore the ease of training our network and its seamless applicability to different datasets. Remarkably, the average precision of the segmentation masks approaches an impressive 98.5% on our exclusive datasets.