Lightweight Algorithm Optimization of Visual SLAM on Embedded Computing Devices for Intelligent Transportation Systems

In the context of intelligent transportation vehicle perception, embedded computing devices serve as the primary computing platform, facing the challenge of the traditional visual SLAM(Simultaneous Localization and Mapping) framework's high computational demands for environmental feature points. To address issues such as point cloud drift errors in long-term, large-scale road traffic perception tasks and the high mismatch rate of feature point tracking in traffic scenes with numerous dynamic objects, this work proposes an optimized feature point mismatch elimination method for the visual odometry module based on the ORB-SLAM3 framework. Additionally, an efficient visual vector dictionary loading and matching algorithm for repetitive keyframes is designed for the loop closure detection module. In the feature point mismatch elimination calculation of the visual odometry module, a feature confidence index is introduced to eliminate mismatched feature points of dynamic traffic objects. Meanwhile, in the loop closure detection module, a binary loading method is applied to optimize visual vocabulary matching, addressing the scene re-localization problem during large-scale road environment loop detection. The work was tested and evaluated on the KITTI dataset using the RDK X3 embedded computing device. Results indicate that the optimized algorithm framework maintains the accuracy of map point coordinate calculations without degradation and achieves real-time performance in large-scale road environment perception and reconstruction tasks in dynamic traffic scenes. Moreover, the computational speed and memory usage of the framework are superior to the original SLAM framework.