KOM-SLAM: A GNN-Based Tightly Coupled SLAM and Multi-Object Tracking Framework.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Simultaneous localization and mapping (SLAM) and multi-object tracking are crucial for autonomous systems.
• Existing methods often process keypoint and object associations separately, limiting performance in dynamic environments.

Purpose of the Study:

• To develop a tightly coupled SLAM and multi-object tracking framework for enhanced robustness in complex dynamic scenes.
• To jointly learn keypoint and object associations across frames within a unified framework.

Main Methods:

• Proposed KOM-SLAM, a novel framework utilizing a Graph Neural Network (GNN) for integrated SLAM and multi-object tracking.
• Constructed a spatiotemporal graph for keypoint and object association, incorporating a multilayer perceptron (MLP) for adaptive thresholding.
• Implemented soft assignment for differentiable pose estimation, allowing direct supervision of association learning via pose loss.

Main Results:

• Demonstrated improved performance on the KITTI Tracking benchmark.
• Achieved superior results in both localization accuracy and object tracking capabilities compared to existing methods.
• Showcased the effectiveness of joint learning for keypoint and object associations in dynamic scenarios.

Conclusions:

• KOM-SLAM offers a robust and effective solution for coupled SLAM and multi-object tracking.
• The GNN-based approach successfully addresses limitations of separate association strategies in dynamic scenes.
• The framework enables direct supervision of association learning through differentiable pose estimation, enhancing overall system performance.