Automated video surveillance systems (AVSs) have recently become vital for ensuring public safety, particularly at events with huge audiences like sporting events. Machine (ML) and deep learning (DL) open the way for computers to think like humans even further by including training and learning components, which artificial intelligence (AI) already provides. In order to evaluate and make sense of surveillance data acquired by fixed or mobile cameras mounted indoors or outdoors, DL algorithms require data labelling and high-performance processors. Recent advances in generative adversarial networks (GANs) for image synthesis and creation in VSSs have made it a hot topic in the field of study to establish if a given input is typical or atypical. Therefore, this research presents a better GAN network to recognise human gaits and to distinguish between human actions that are normal and pathological in VSSs. To achieve this goal, we first combine global and local features to enhance learning in crucial local regions that include multiple key points. Two, we use metric learning to pull out shared and unique characteristics. After features have been retrieved, they are used as input by the classification module in order to identify GAN-generated pictures. DMO is used in this study to perform hyper-parameter optimization (HPO) in GAN, which provides significant metaheuristic balance between the survey and misuse phases.