Sarcouncil Journal of Multidisciplinary

Abstract: The abundance of new technologies regarding intelligent surveillance systems has contributed to the emergence of the so-called efficient hybrid deep learning pipelines, according to which an anomaly may be detected in real time, to be more precise, the anomaly that will result in the appearance of an abandoned object in the crowd or a person in the setting. These systems integrate object recognition, tracking, and temporal behavior analysis to give rise to a better response to security and situational awareness. The article provides a summary of the new developments in the neural networks of deep learning (CNNs), long short-term memory (LSTM) units, attention models, and object re-identification units. A comparison of such models as You Only Look Once (YOLO) and Simple Online and Realtime Tracking (DeepSORT) is made against each other in terms of real-time performance, computational and contextual intelligence speed. Among the main focuses are the relationship between the objects and the individuals, the deployment of edges, the identification of small objects, and the future research topics such as the diversity of objects, the problem of model interpretability, and ethics. The present review has given an overview of what is being done and suggests that the transformational power of the hybrid architecture can be employed to complement real-time surveillance systems.