Semantic SLAM using laser-vision data fusion: Enhancing autonomous navigation in unstructured environments

The increasing complexity of autonomous navigation environments has necessitated the integration of multi-sensor data into Simultaneous Localisation and Mapping (SLAM) algorithms, as reliance on single-sensor data has proven insufficient. Recent research has focused on utilising fused data from multiple sensors to enhance SLAM performance, particularly in creating semantic maps that combine high storage efficiency with low ambiguity and enriched information content. Semantic maps, which embed semantic information into spatial representations, enable robots to better interpret complex environments and improve their autonomous operation in unstructured, GPS-denied, and extreme settings. In this context, an advanced semantic SLAM algorithm, termed NN-SLAM, is proposed based on the fusion of visual and laser data. For semantic segmentation, the PSPNet_CBAM network framework was developed, achieving improvements of 0.73 %, 0.14 %, and 0.92 % in mean Accuracy (mAcc), Pixel Accuracy (PAcc), and mean Intersection over Union (mIoU), respectively, on the Cityscapes dataset compared to the original PSPNet model. For the fusion of visual and laser information, NN-SLAM demonstrated a reduction in filter processing time by 0.2440 ms and 3.6497 ms on the KITTI 0011 and 0093 datasets, respectively, when compared to the Visual-Lidar Odometry and Mapping (ALOAM) algorithm. Additionally, the size error between the generated map and the actual scene was reduced to just 0.81 %, indicating high fidelity in mapping accuracy. In terms of semantic mapping, the region-growing algorithm was utilised to optimise the semantic mapping and data association process, yielding superior performance in semantic projection tasks. These results highlight the potential of NN-SLAM in advancing the field of autonomous robotics by providing robust, efficient, and accurate solutions for navigation in complex and challenging environments.