Metaheuristic-Tuned GraMNet Architecture for Enhanced Video-Based Anomaly Detection using UCF50 Dataset

Abnormal activity detection in video surveillance has become a critical research domain, driven by the increasing demand for intelligent security systems in public and private environments. Abnormal activity recognition involves identifying unusual or deviant behaviors within a monitored environment, often using video surveillance, sensor data, or wearable devices. This process is crucial in requests such as security monitoring, healthcare, besides smart homes, where detecting anomalies like falls, unauthorized access, or erratic movements can prevent harm or trigger timely interventions. By leveraging deep learning models besides spatio-temporal analysis, systems can learn typical behavior patterns and flag activities that diverge from norm. This study presents a novel hybrid deep learning framework integrating GramNet (GraMNet) architecture with bio-inspired Greater Cane Rat Algorithm (GCRA) for hyperparameter tuning. UCF50 dataset, comprising 50 real-world action categories, was employed to train and validate proposed model. GraMNet architecture utilizes modular SubNets conFig.d in both serial and parallel forms, enabling efficient feature extraction while maintaining computational feasibility. GCRA, inspired by intelligent territorial and foraging behavior of Greater Cane Rats, was applied for hyperparameter optimization to improve convergence and model generalization. system incorporates transfer learning to mitigate challenges of data scarcity and boosts classification accuracy. Experimental evaluations demonstrate that proposed GraMNet-GCRA framework achieves superior performance across multiple metrics, with a peak accuracy of 96%, precision of 94%, recall of 93%, besides F1-score of 93.5%, outperforming state-of-the-art metaheuristics such as AOA, WOA, DMO, and ADMO. Moreover, it significantly reduces training time without compromising model efficacy. Visual analysis using ROC and loss curves confirms stable training dynamics and effective anomaly classification. This work highlights potential of combining structured CNN-based models with biologically-inspired optimization for enhanced abnormal activity recognition in video data.