AHSWN Home · Issue Contents · Forthcoming Papers

Low-Pixel Panoramic Surveillance Synthesis Based on LightGlue
Xiaofan Pang, Kaiwen Zheng, Chengze Du and Zhuang Shi

Panoramic monitoring, widely used in security, virtual reality, and robotics, combines images from multiple cameras to create seamless views. However, traditional stitching methods like SIFT and SURF struggle with low-resolution monitoring images, leading to feature loss, image dislocation, and slow processing. We propose a novel stitching method based on L-LightGlue to address these limitations, incorporating multi-scale feature fusion and image hyper-segmentation. This approach enhances precision and robustness by reducing feature loss and accelerates processing by matching feature points across scales. The architecture leverages an attention mechanism for efficient sparse feature matching, significantly improving the speed and accuracy of stitching low-pixel images. Experiments demonstrate that our method surpasses traditional techniques in real-time performance, stability, and accuracy. This advancement makes panoramic monitoring more reliable and scalable, broadening its applications in public safety, traffic oversight, industrial operations, and urban planning, even with hardware and resolution limitations.

Keywords: Image stitching, Feature matching, Deep learning, Attention mechanism