Enhancing security systems: human activity recognition using transfer learning model

Enhancing security systems: human activity recognition using transfer learning model

연구 분야: Infrastructure

학회: International Journal of Information Technology

In today’s environment, human activity recognition through video surveillance is essential for improving security by threat detection, prevention, and reporting. It makes automated video data surveillance possible with applications: proactive risk assessment in the industrial setting, suspicious behavior detection in smart home environments, and proactive threat detection in public spaces by detecting suspicious behavior. Deep learning algorithms provided by artificial intelligence (AI) offer automatic video monitoring by evaluating unusual, unapproved, or dangerous circumstances and allowing prompt action. The proposed article explores the performance of proposed hybrid deep learning models: VGG16-GRU, VGG19-GRU, and DenseNet121-GRU, on the UCF-11 dataset, which consists of 11 different categories of real-world activity. The proposed architectures include convolutional neural networks (CNNs) based spatial feature extraction while gated recurrent units (GRUs) explore the temporal features of video sequences. The DenseNet-GRU model uses the dense connection to improve feature reuse and representation learning. The features extracted by the VGG16-GRU model are more effective and portable. According to the experimental results, the DenseNet121-GRU is the best performing hybrid HAR model with an accuracy of 93.5%. The notable highlight in the proposed study is the use case addressing the enhanced security systems with threat detection, prevention, and reporting. The proposed study underscores the potential of hybrid CNN-GRU architectures for developing scalable and effective HAR solutions, particularly in advanced security systems.