Suspicious human activity detection model via improved deep joint segmentation and hybrid model

In general, suspicious human activities do not occur in front of the cameras, making it more challenging to detect the activities via Closed Circuit Television (CCTV) footage. The problem of detecting suspicious activity in low-quality videos is not sufficiently addressed by state-of-the-art efforts. Unlike traditional models that rely on limited spatial or temporal features, the proposed framework introduces a multi-level, deeply integrated approach that enhances both feature representation and behavior classification. The proposed model has several innovations such as enhanced preprocessing, advanced feature extraction, hybrid deep learning architecture and improved score-level fusion. First, during the pre-processing phase, the video is altered to frames. Pre-processing is then carried out after enhanced Deep Joint (DJ) segmentation. Next comes feature extraction, which involves drawing the Shape Local Binary Texture (SLBT), enhanced Local Gradient Increasing Pattern (LGIP), and Hierarchy of Skeleton (HOS) features. The core Artificial Intelligence (AI) implementation utilizes Local Binary Pattern (LBP)-embedded Convolutional Neural Networks (CNNs) in integration with Long Short-Term Memory (LSTM) networks to effectively model and recognize temporal behavior patterns. Finally, the proposed work adopts improved Score level fusion for getting the ultimate result of suspicious behaviors. The application of AI in this work revolves around the use of deep learning models (CNNs and LSTMs) to analyze video data, extract meaningful patterns, and ultimately identify behaviors that deviate from the norm, making it a powerful tool for automated surveillance systems. The proposed method offers significant advantages with more reliable suspicious behavior identification through the fusion of spatial and temporal models.