Crowd counting in intelligent video surveillance systems

This study focuses on enhancing the accuracy and robustness of crowd counting in intelligent video surveillance systems by incorporating perspective-awareness into deep learning models. Traditional convolutional neural networks often struggle with scale variations caused by perspective distortions and object occlusions in dense crowd scenes. The goal of this study is to develop a method that leverages geometric depth information to improve the spatial consistency of density estimations and provide more reliable predictions across varying scene configurations, including highly congested or irregular environments. The tasks to be accomplished include designing a depth inclusion module, integrating it into an encoder-decoder architecture, generating depth maps from monocular RGB images, and recalibrating feature representations using attention-weighted scale-aware mechanisms. The methods used involve the extraction of depth features from images via a pre-trained depth estimation model, followed by spatial attention-based recalibration of feature maps to highlight foreground objects and suppress irrelevant background signals. A fully differentiable pipeline is implemented to ensure seamless integration into standard CNN frameworks. The network training procedure also incorporates Euclidean loss functions on pixel-level density maps to optimize scale-sensitive prediction. The proposed method is evaluated on benchmark datasets including ShanghaiTech-B, UCF_CC_50, and Mall, where it consistently outperforms state-of-the-art models in terms of MAE and MSE. The experimental results confirm that the explicit incorporation of depth-aware representations significantly enhances counting performance, especially in scenarios with severe perspective-induced scale disparities. The integration of geometric priors into data-driven models offers a promising direction for real-time surveillance and large-scale crowd monitoring applications, providing not only quantitative improvements but also greater spatial fidelity in density map generation and better adaptability to complex visual conditions.