基于行人检测与闭环阈值的二维激光 SLAM 方法

Objective In recent years, service robots have been widely deployed in indoor environments. Their reliable operation fundamentally depends on environmental mapping and self^-localization, known as simultaneous localization and mapping (SLAM). Due to its wide field of view, high accuracy, and strong environmental adaptability, two^-dimensional (2D) lidar is commonly employed for environmental perception in indoor robot SLAM. However, the inherently limited scanning range of 2D lidar renders 2D laser SLAM systems vulnerable to disturbances from dynamic objects. Passing pedestrians or moving furniture, such as chairs and tables, may introduce instability into SLAM, generating erroneous features or ghost artifacts in the constructed map, causing local map^-matching failures due to localization errors, and resulting in incomplete map alignment during loop closure. To address the challenges of localization failure and degraded map quality in dynamic indoor environments, this paper proposes a 2D laser SLAM method based on pedestrian removal and adaptive^-threshold graph optimization, thereby enhancing both mapping and localization performance in dynamic scenario. Methods The method comprises two stages: front^-end pedestrian filtering and back^-end map optimization. In the front^-end pedestrian filtering stage, an improved adaptive clustering algorithm segments the raw point cloud, eliminating stray noise points and refining the clustering results. The raw point cloud is then down^-sampled by 1/2 and 1/3 to generate a multi^-scale representation with three channels. Because laser echo intensity depends on target surface properties—such as material composition, surface roughness, and incident angle—geometric features are fused with intensity information to form geometric-intensity multi^-modal point cloud inputs. Subsequently, a multi^-branch convolutional neural network (CNN) architecture processes these multi^-scale and multi^-modal features, enabling real^-time pedestrian detection and precise removal of the corresponding point clouds. In the back^-end optimization phase, an adaptive threshold graph optimization strategy reduces unnecessary matching constraints by applying matching node thresholds, thereby enhancing computational efficiency during graph optimization. Results and Discussions To evaluate the performance of the pedestrian detection model, a leg dataset is collected from point clouds gathered in two scenarios: public halls and rooms. Three ablation studies are conducted to assess the effects of feature modality, point cloud scale, and CNN branch configuration. Quantitative metrics including accuracy, precision, recall, and F1^-score are employed. Results indicate that recognition accuracy increases with higher laser intensity (Table 2), higher point cloud resolution (Table 4), and more CNN branches (Table 5). The proposed model, which integrates multi^-modal features, multi^-scale representations, and a multibranch CNN architecture, achieves 96.38% recognition accuracy, 5.63 percentage points higher than the random forest method. To validate pedestrian removal effectiveness, point cloud elimination tests are performed in a dynamic commercial building hall. Experimental results demonstrate that the proposed algorithm effectively detects and removes pedestrian points during mapping (Fig. 16). To assess adaptive^-threshold graph optimization, distance (0.5 m) and angle (0.05 rad) thresholds are used as constraints. With 28 sub^-maps in the matching graph, the adaptive threshold reduces the total number of matches by 80% [Fig. 17(b)] and decreases optimization time from 1140.98 μs to 227.37 μs [Fig. 17(a)], significantly reducing computational load. By integrating frontend pedestrian point cloud filtering and back^-end adaptive threshold graph optimization, the point clouds are processed and fed into the improved Cartographer to generate a map. The resulting map exhibits reduced ghost artifacts and improved loop closure performance (Fig. 18), confirming the robustness of the proposed 2D SLAM method in dynamic indoor environments. Conclusions To mitigate pedestrian^-induced interference in dynamic SLAM scenarios, this paper proposes a 2D laser SLAM method that integrates pedestrian detection with adaptive loop^-closure thresholds, refining both the front^-end and back^-end of the SLAM pipeline. The front^-end employs adaptive thresholding to eliminate stray noise points, expands multi^-dimensional and multiscale point cloud features, and achieves real^-time detection and removal of pedestrian point clouds through CNNs with 96.38% accuracy. A graph optimization algorithm incorporating loop^-closure thresholds is introduced to reduce the computational complexity of matching operations, significantly enhancing map consistency and coherence. Validation tests in a commercial building hall confirm the mapping robustness of the proposed algorithm in dynamic environments.