Efficient SLAM Approach with MultipleCameras Based on RGB-D Setup

Kurzfassung

For autonomous navigation in unknown environments, a robust and continuous state estimation is extremely valuable. It builds the foundation for most vision-based applications such as robotics or autonomous cars. Due to their light weight and small size, cameras are often utilized as sensors on robotic platforms. They offer a broad spectrum of applications, such as simultaneous localization and mapping (SLAM), 3D reconstruction or object recognition. Due to the fact that cameras can be used in a diverse range of areas and are inexpensive to purchase, robot platforms are often equipped with multiple cameras. For the accurate state estimation, five RGB-D cameras are available to the robot used in this work. While most of the research work in the thriving field of SLAM is based on the use of only one camera, in this work a multi-camera SLAM system is proposed that can incorporate any number of RGB-D cameras. For this purpose, a feature-based approach is chosen. The camera images are reduced to a sparse set of image features to enable fast processing and guarantee state estimation at high frame rates. At the same time, a sparse map of the surrounding area is constructed. Among other purposes, the robot then uses this map to improve its state estimating. To test the multi-camera SLAM approach, a dedicated dataset is recorded that consists of six scenarios and corresponding ground truth trajectories. It demonstrates its correct operation. The time analysis indicates the bottlenecks when adding more cameras. The quantitative evaluation results show that the multi-camera approach outperforms state-of-the-art SLAM systems in terms of both accuracy and robustness.