Multimedia - A Multi-Focal Dynamic SLAM

In order to achieve wider perceptual field of view, longer measurement distance, and higher positioning accuracy, a multi-focal dynamic visual simultaneous localization and mapping (SLAM) is proposed in this paper. The definition of multi-focal stereo and the camera calibration procedure are first presented. The number of extracted and matched features will decrease as a result of the big scale difference of multi-focal images, which will impair the positioning accuracy of visual SLAM (VSLAM). Then, an effective feature extraction and matching method based on adaptive image pyramids is given, and the matched features are triangulated to obtain their depth. The traditional VSLAM approaches are predicated on the fundamental premise that the environment is static, which causes the positioning and mapping accuracy being completely unreliable in a dynamic environment. A deep learning segmentation method is adopted to obtain priori dynamic objects, while multi-view geometry, regional feature flow, and inverse perspective mapping are combined to verify dynamic objects and eliminate dynamic features. These methods are integrated with multi-focal cameras to realize multi-focal dynamic SLAM (MF-DynaSLAM). MF-DynaSLAM is compatible with both same-focal and multi-focal stereo, and the performance of the proposed method are evaluated on KITTI benchmark and self-built datasets. The suggested adaptive image pyramids can boost the number of matched features by 85.36% when the focal length ratio of two images is 2.1. Comparison to ORB-SLAM3 and DynaSLAM, our method has lower positioning absolute trajectory error (ATE). This approach is also real-time and operates at a processing rate of about 10FPS.