An improved multi-object classification algorithm for visual SLAM under dynamic environment

In this paper, a novel dynamic multi-object classification method is proposed based on real-time localization of a robot equipped with a vision sensor that captures a sparse three-dimensional environment. Specifically, we build upon ORB-SLAM2 and improve its formulation to better handle moving objects in a dynamic environment. We propose a feature classification algorithm for ORB (oriented FAST and rotated BRIEF) features in complex environments. Based on inter-frame texture constraints, we add the reprojection error algorithm, which can reduce the influence of illumination and dynamic objects on the simultaneous localization and mapping (SLAM) algorithm. We then propose a new dynamic initialization strategy and apply the proposed feature classification algorithm to the ORB-SLAM2 tracking thread. For real-world implementations, we focus on the robustness and real-time performance of dynamic target segmentation simultaneously, which cannot be satisfied by existing geometric segmentation and semantic segmentation methods. From an engineering point of view, the proposed work can quickly separate dynamic feature points based on traditional methods, which makes the proposed algorithm has better real time in practical applications. We thoroughly evaluate our approach on the TUM and KITTI benchmark database and on a real environment using the Turtlebot platform equipped with a Bumblebee camera. The experimental results indicate that the proposed method is more robust and accurate than current state-of-the-art methods in different environments.