Deep Learning Based Dynamic Object Filtering for Semantic Information Assisted Visual LiDAR SLAM in Dynamic Scenes

To address the structural mismatch between semantic priors and motion evidence in tightly coupled visual-LiDAR SLAM, this paper proposes a dynamic object filtering architecture centered on evidence arbitration. Semantic segmentation outputs are transformed from binary rejection masks into continuous spatial modulation coefficients ranging from 0 to 1. These coefficients guide the generation of two complementary streams of motion evidence: visual optical flow residuals and multi-frame LiDAR occupancy consistency. Log-odds fusion then produces cross-modal confidence and its uncertainty. Using semantic instances as arbitration units, the method resolves conflicts within a sliding window of 6-8 frames by jointly considering mean confidence, temporal stability, and spatial concentration, with an upper decision delay bound of 3 frames. Finally, dynamic probabilities are embedded in the factor graph as information matrix modulation factors, replacing direct front-end rejection. Experiments on Semantic KITTI, KITTI raw, and Urban Nav show that, while maintaining a dynamic-point residual rate no higher than that of the hard-filtering baseline, the method increases the static semantic constraint retention rate from 0.42 to 0.78. Decision accuracy reaches 0.86 and 0.83 for conflict modes (b) and (c), respectively, and trajectory errors during transition intervals are substantially lower than those of the comparison variants. A breakdown of per-frame processing time identifies semantic segmentation as the main bottleneck. The system meets real-time constraints when the segmentation frame rate is at least 8 Hz, and the dynamic instance density does not exceed 15 instances per frame.