Hidden Object Detection Using Enhanced Transformer With Spatio‐Temporal Feature Extraction

ABSTRACT

Real‐time video surveillance faces difficulties during hidden object detection because of various factors, which include occlusions and changes in lighting, and moving background elements. The research presents a new hybrid deep‐learning system that enables effective hidden object detection from live video streams that public‐area surveillance networks capture. Users extract key frames from input video sequences, which enables them to decrease redundancy while maintaining essential motion details. The EffDet‐GRU model uses its combination of EfficientDet and gated recurrent units (GRU) to perform feature extraction through spatial–temporal dependency capture. The enhanced transformer architecture processes these features to establish hidden‐object localisation through its long‐range contextual relationship modelling. The chaotic sinusoidal mapping‐based hiking optimisation (CSM‐HO) algorithm functions as a fine‐tuning tool that enhances convergence while reducing local minima problems because it optimises the transformer's loss function. The abandoned objects dataset testing results show that our system outperforms all other systems because it achieves 99.46% accuracy, 99% precision, 98.94% recall and 99.08% F1‐score. The framework delivers reliable hidden object detection results that operate almost in real‐time to support security and surveillance system needs.