Diffusion Model-Based Pedestrian De-Occlusion Method Using a Monocular Camera Sensor for Indoor Visual Localization

In monocular camera-based visual localization methods, query images are compared with reference images in the database using an image retrieval algorithm. Once the best matching image is obtained, the pose of online users can be estimated using algorithms such as epipolar geometry or EpnP. However, in practical applications, the quality of query images may be affected by environmental and device factors, resulting in decreased image quality. This article primarily focuses on addressing the challenges posed by pedestrians, which are commonly encountered obstacles in indoor visual positioning scenes. It proposes an automatic approach for performing semantic segmentation and occlusion removal to enhance query image quality and improve positioning accuracy. Extensive experimental analysis and comparisons of positioning errors demonstrate that the proposed method significantly enhances the performance of visual localization techniques. To the best of our knowledge, this is the first attempt to address occlusion interference in vision-based localization through an active removal approach.