Abstract:
Based on the rapid development of deep learning, semantic information has gradually become a research hotspot in the field of SLAM (Simultaneous Location and Mapping). Due to the noise problem caused by the environment and the sensor, there are some abnormal points in the semantic map constructed by most existing semantic SLAM algorithms, which results in the lack of consistency of the constructed semantic maps and affect the accuracy of the algorithms. The loss function can adjust the weights assigned to the outliers, thereby suppressing the existence of the outliers. However, the model of loss function used by most semantic SLAM algorithms is fixed and cannot adapt well to changes in ambient noise. To solve this problem, a robust lidar semantic SLAM algorithm with variable structure is proposed, which uses Gaussian mixture correntropy weight function as the loss function, called VS-SLAM. It uses the characteristics of the model structure that can be adjusted by adjusting the parameters and changed with the surrounding environmental noise to fit the noise distribution to the greatest extent, which is more conducive to reducing the weight of the algorithm for outliers and improving robustness to the outliers. Experiments on the open KITTI dataset show that the average relative translation error and rotation error are reduced by 5.36% and 8.82% respectively, while the mapping time is almost equal, and the constructed semantic maps are more consistent.