SLAM Algorithm Based on Fusion of Improved Point and Line Features with Inertial Sensing Units

In response to the issue of excessive line segment extraction by traditional simultaneous localization and mapping (SLAM) algorithms in dense environments, a SLAM algorithm based on improved point line features and tightly coupled inertial sensing units (IMU) was proposed, namely IPLI-SLAM. During the data preprocessing phase, Shi-Tomasi features were used for point feature extraction, and the Lucas-Kanade (LK) optical flow was used for tracking and matching. Line features were introduced to improve the line segment detector (LSD) extraction algorithm based on pixel gradient filtering mechanism, and to filter out areas with dense line features. By tightly coupling visual point and line information with IMU and adding it to the backend, the algorithm accuracy was improved. Finally, IPLI-SLAM algorithm was tested and validated on the dataset EuRoc and in actual scenes.The results show that compared with the original LSD algorithm, the improved LSD algorithm reduces extraction and matching time by 8.2%. Compared with the VINS-mono and PL-VINS algorithms, the IPLI-SLAM algorithm has improved positioning accuracy by 50.7% and 13.2%, respectively. In environments with fast motion speed and blurred scenes, the positioning accuracy is further enhanced by 55.6% and 25.1%, respectively. In actual multi-texture scenes, the distance between the starting and ending points of the localization trajectory using our algorithm is smaller than that of VINS-mono algorithm. This proves that the algorithm proposed in this article not only significantly improves positioning accuracy, but also has higher stability and stronger robustness.