Візуально-інерційний SLAM з використанням розширеного фільтра Калмана для автономної навігації

Abstract

The study is devoted to the use of visual-inertial SLAM based on the extended Kalman filter (EKF) for autonomous navigation tasks. The study used data from the KITTI dataset, which include images from a stereo camera, as well as linear and angular velocities obtained from an inertial measurement device. To determine visual landmarks, it is proposed to use ORB features, which are characterized by their speed of calculation and resistance to changes in lighting, rotation and scaling. In addition, an algorithm for selecting relevant landmarks has been developed, which allows to increase the accuracy and speed of SLAM. The results of a series of experiments have shown that the efficiency of the system largely depends on the adjustment of the noise ratio of the motion model (Q) and sensor measurements (R). It was determined that the optimal range of Q/R for the studied datasets is 0.001…0.00001. To achieve a balance between speed and system performance, the influence of the number of relevant visual features in the range from 10 to 120 was investigated. The results of the experiments showed that the optimal number of landmarks is 50…80, and the best results were achieved at Q/R=0.001 or Q/R=0.00001, depending on the dataset used. The trajectory accuracy was assessed using the ATE metric (Absolute Trajectory Error). GPS data was used as ground truth for verification. The results confirmed that visual-inertial SLAM is an effective tool for autonomous navigation, especially in the absence of GPS. The use of EKF in visual-inertial SLAM reduces computational costs compared to other methods, which makes it optimal for devices with limited resources, in particular mobile robots. Further research will focus on optimizing landmark selection, improving noise modeling, and adapting the algorithm for monocular cameras to improve SLAM accuracy and efficiency. Further research will focus on optimizing landmark selection, improving noise modeling, and adapting the algorithm for monocular cameras to improve SLAM accuracy and efficiency. In addition, research on the invariant extended Kalman filter (IEKF) is planned, which has the potential to improve mapping accuracy and enhance system robustness in complex and dynamic environments.