An enhanced artificial learning framework for moving object detection in challenging video scenes

A visual surveillance system plays a critical role in enhancing public safety by enabling real-time monitoring and recording of activities in public spaces. Change detection followed by object tracking are fundamental steps in such systems. However, detecting objects moving at different speeds poses significant challenges. Fast-moving objects often result in motion blur, hindering detail capture, while slow-moving objects may be misclassified as static, especially in cluttered environments. Additionally, occlusion and overlapping objects complicate tracking in crowded scenes, and system performance decreases in unseen video setups. Traditional approaches frequently struggle with these issues. This research introduces a novel approach to address these limitations, emphasizing the importance of background subtraction (BGS) for change detection. A new encoder–decoder architecture is proposed for detecting moving objects in complex video environments. The encoder leverages a modified VGG-16 network, enhancing accuracy by extracting fine- and coarse-scale features crucial for precise detection. The proposed feature extraction framework (FEF) combines average pooling, convolutional layers, and convolutional layers with varying sampling rates to achieve multi-scale and multi-dimensional feature representation. The decoder, built with stacked transposed convolution layers, contrast normalization (CN) layers, and LeakyReLU (LR) activation functions, efficiently reconstructs these features into the image space. Rigorous testing against 46 existing methods demonstrates the system’s superior performance. Both subjective and objective evaluations validate its effectiveness, while its robustness is further highlighted in unseen video setups, showcasing significant advancements in visual surveillance technology.