Human Detection in Low-Resolution Depth Images

Human detection is essential in areas like security surveillance, healthcare, and behavior analysis. While traditional RGB-based methods are effective, they raise significant privacy concerns by exposing personal identities. Depth images, in contrast, provide a privacy-preserving alternative but are often hindered by the limited availability of low-resolution datasets. Despite their potential, human detection in low-resolution depth images remains an underexplored area, primarily due to the challenges posed by poor spatial resolution and insufficient depth data. These images often lack sufficient detail, leading to blurred human contours and making detection more difficult. To address this gap, we design and construct a novel low-resolution depth human detection dataset, offering a rich resource of labeled data. To tackle the issue of insufficient detail in low-resolution depth images, we introduce a spatial perception network module that captures and leverages spatial relationships between regions. This module significantly enhances detection accuracy by improving the model's understanding of spatial context. Our proposed method offers a promising solution for real-world applications where privacy and low-resolution data are key considerations.