A moving object classification and dense sampling method for dynamic object tracking VSLAM system

Purpose Learning-based methods fail to use geometric information fully and are vulnerable to segmentation accuracy. This paper introduces a feature point classification module to the tracking thread, which uses a Gaussian distribution to model the interframe distance variation induced by combining scene flow and epipolar constraint. This study aims to improve the positioning accuracy by determining the motion state of the object, which can accurately obtain the attitude and state information of the object. Meanwhile, the authors designed an adaptive sampling strategy based on the distance transformation, aiming to in order to reduce the pose estimation error of dynamic objects during tracking. Design/methodology/approach The algorithm introduces a Gaussian distribution to model interframe distance variations and combines scene flow and epipolar constraints to perceive the motion state of real objects. In addition, objects recognized as dynamic objects are used to design an adaptive dense sampling strategy based on distance transformations to filter noise. Finally, this paper evaluates the proposed algorithm on the publicly available KITTI dataset and perform a validation using a mobile robot. Findings This paper evaluates the proposed algorithm on the publicly available KITTI dataset and validates it using a mobile robot. The experimental results show that the proposed algorithm is closely related to the real situation and can accurately estimate the attitude of dynamic objects. Social implications This approach provides new ideas for combining deep learning with geometric methods and provides a path to expand the traditional methods to remove noise points. Originality/value The algorithm provides a new idea for the combination of deep learning and geometric methods with applicable localization accuracy, better estimation of object attitude, extends the traditional method of removing noise points, and is crucial for improving the autonomous navigation ability of robots in dynamic environments, which is also of great significance for applications in the field of industrial robotics.