A dynamic obstacle culling 3D real-time map building algorithm for mobile robot localization

In this study, the complexity of an ICU environment in terms of both static and dynamic objects when building a 3D real-time map for mobile robot localization is solved. YOLOv5, probability dynamic feature points culling algorithms and ORB-SLAM3 are fused and dynamic objects are culled using a probability based culling rule which eliminates the dynamic objects and retains the static ones. The algorithm utilizes a point cloud construction thread to establish a dense point cloud map from the sparse point cloud results output through a point cloud stitching thread after the elimination of the dynamic feature points and this is converted and saved as octree map for memory efficiency. Evaluation of the accuracy of the proposed 3D real-time map building algorithm performed using the f3/walking_xyz dataset from the TUM RGB-D dataset sequence utilizes the relative pose error (RPE) and absolute trajectory error (ATE) as metrics for dynamic obstacle culling and mapping accuracy. For a dynamic feature point elimination experiment involving 8 keyframes. The study reveals that 28 dynamic feature points remain in 8 keyframes, with a success rate of 89.74% for dynamic feature point elimination. The remaining feature points are mainly in "chair" type objects. The algorithm, based on YOLOv5s-PROBABILITY CULLING-ORB-SLAM3 fusion effectively eliminates dynamic objects, with low false rejection rates, optimized feature point extraction, and significantly improves mapping effect and accuracy by 91.4% compared to ORB-SLAM3 implemented alone which yields 72% in environment like ICU, where frequent movement is prevalent, demonstrating its superior accuracy, effectiveness and precision.