ITM-SLAM: Instance-Level Semantic Tracking and Mapping in Dynamic Scenes

Simultaneous Localization and Mapping (SLAM) constitutes a critical component of mobile intelligent robotics. However, it often faces challenges due to the influence of dynamic objects in many scenarios. Consequently, to address these challenges, this paper introduces an instance-level semantic SLAM algorithm, ITM-SLAM. It advances in visual odometry, data association, and nonlinear optimization to effectively handle dynamic objects. Initially, a visual odometry based on object detection and point cloud clustering is designed, alongside edge detection for accurately delineating the boundaries of dynamic objects. Moreover, to estimate the motion of dynamic objects, the algorithm employs the Kalman filter. Additionally, it designs a strategy for object allocation to facilitate data association. In terms of optimization, this paper proposes an optimization method based on ellipsoid and dynamic feature, integrating dynamic objects into the overall optimization framework. Experimental evaluation conducted on the TUM dataset demonstrates that the algorithm proposed in this paper significantly enhances localization accuracy in comparison with various SLAM algorithms.