A Novel Deep Learner for Human Behavior Prediction Over Public Video Surveillance

Identifying human behavior effectively is essential for spotting anomalies in video surveillance systems, particularly in dynamic environments. Conventional methods frequently have significant false detection rates, which restricts their use. This paper proposes a unique framework to improve anomaly classification by combining improved preprocessing, feature extraction, optimization methods, and Remote Field-Based (RFB) video sensors. This work’s primary contribution is the creation of an all-encompassing system that uses feature-driven learning and optimization to minimize false alarms and achieve high detection accuracy. The UBI-Fight and PETS2009 datasets, which are often used in surveillance applications for human behavior detection, are used as validation standards by the framework. Key frame extraction is the first step in the procedure, which is then followed by sophisticated preprocessing methods like illumination correction, occlusion control, and noise reduction. Harris corner detection, Histogram of Oriented Gradients (HOG), and Particle Gradient Motion Analysis (PGMA) are used to extract motion and spatial data. By capturing high-resolution motion dynamics for improved data collection, the integration of RFB video sensors improves motion analysis and behavioral evaluation. While Particle Swarm Optimization (PSO) optimizes the classifiers to increase accuracy, a deep learning classification strategy is used for anomaly classification. The suggested methodology outperforms conventional techniques, achieving 93.4% accuracy on the PETS2009 dataset and 89.2% accuracy on the UBI-Fight dataset. These outcomes show how well feature extraction and optimization work together to improve the recognition of human behavior. Reliable security monitoring is ensured by the framework’s scalability and real-time applicability, which also considerably reduce false alarms and offer strong performance in a variety of public safety scenarios.