Exploration of SLAM Technology Based on the Fusion of Stereo Cameras and IMUs

With the rapid advancement of science and technology, artificial intelligence has garnered significant attention. Simultaneous Localization and Mapping (SLAM)—a critical technology for mobile robotics—has seen substantial development. However, as application scenarios become increasingly diverse and complex, vision-only SLAM systems often suffer from reduced localization accuracy and insufficient robustness when faced with challenges such as rapid motion, a lack of visual texture, or fluctuating lighting conditions. By integrating an Inertial Measurement Unit (IMU)—which operates independently of environmental constraints and provides reliable motion data—we can explore SLAM techniques that fuse stereo camera data with IMU inputs, thereby enhancing the performance of SLAM systems in both localization and mapping.
(1) To address the challenge of pose and state estimation in SLAM systems, this study analyzes the underlying mathematical models, structural principles, and sources of error. A unified state estimation model—encompassing camera poses, IMU states, and landmark positions—was constructed and solved. Furthermore, to mitigate the impact of lens-induced distortions on visual measurement accuracy, the system accounts for both radial and tangential distortions. Through a comprehensive calibration process, key parameters—including intrinsic camera parameters, distortion coefficients, and the extrinsic parameters defining the spatial relationship between the stereo camera and the IMU—were determined. Experimental results indicate that the re-projection errors for both the left and right cameras remain within a range of ±0.4 pixels, confirming that the calibration precision meets the operational requirements of the SLAM system.
(2) Within the visual odometry module, the traditional ORB algorithm often suffers from uneven feature point distribution and suboptimal matching performance due to fixed thresholds. To overcome these limitations, this study introduces an improved approach featuring key-region selection and adaptive threshold detection, while optimizing the logic for feature point distribution and extraction. Experimental validation demonstrates that the time required for FAST feature extraction is reduced by 23.3% compared to the traditional algorithm. Moreover, the improved ORB algorithm exhibits superior matching accuracy across various experimental conditions with differing lighting levels. (3) A tightly-coupled fusion framework was developed for processing data from cameras and an IMU. First, the IMU measurement model and integration principles were analyzed, and IMU calibration was performed. Next, the fusion of the stereo camera system and the inertial navigation unit was initialized to obtain reliable estimates of the system's initial pose and velocity. Building on this, a sliding-window-based optimization approach was implemented for the back-end. State variables—categorized into camera-related and non-camera-related components—are processed via marginalization across two separate threads. Optimization objectives were formulated by incorporating both IMU measurement constraints and visual constraints, with a Schur-complement-based marginalization technique employed to accelerate the computation.
(4) To validate the SLAM system, the proposed algorithm was compared against OpenVINS and VINS-Fusion within an Ubuntu 18.04 environment, using the EuRoC dataset. Additionally, data collection experiments were conducted in real-world scenarios following pre-planned paths. Finally, the system was evaluated using the 'evo' tool. The experimental results demonstrate that the proposed algorithm achieves lower maximum error (Max) and root-mean-square error (RMSE) values ​​compared to the other two algorithms, confirming the effectiveness of the proposed improvements.