Real-time Semantic SLAM Mapping Based On Multi-resolution Cascaded Neural Network

In the current development of science and technology,the universal application of artificial intelligence technology has become a status quo,especially the development of robotics has received great attention.Simultaneous localization and mapping(SLAM)is to study the positioning,navigation,and path planning of robots.Adding semantic information to SLAM systems is an important research to improve the humancomputer interaction capabilities of robots.This paper proposes a multi-resolution cascade network for RGB-D input to achieve image semantic segmentation that balances accuracy and speed,and proposes a multi-resolution cascade network-based semantic SLAM mapping module.The multi-resolution cascade network proposed in this paper is for the case of RGB-D multi-dimensional input,which includes image input at three resolutions,and uses depth-similar convolution to fuse depth image information with low-cost computing cost and memory consumption.The feature map obtained through the shallow feature extraction layer at medium resolution is used as a low-resolution branch for further feature extraction.The features of the three resolution branches make the network have a good ability to extract global semantic information and local detailed information.Multi-resolution features are used for fusion,and tags are used for supervised training,and real-time efficiency is achieved in the prediction process.The SLAM semantic mapping module based on the multi-resolution cascade network proposed in this paper can obtain the segmentation result prediction map for RGBD input with real-time efficiency.The SLAM visual odometer module inherits the existing RGB-D-based SLAM system framework,we introduce multi-resolution cascade network in the mapping module to improve efficiency and increase the density of the mapping.We tested and verified image semantic segmentation tasks on multiresolution cascade networks in NYUv2 and SUN RGB-D datasets,and compared with other excellent algorithms.The proposed network structure can better balance accuracy and running speed.At the same time,we implemented the proposed SLAM semantic mapping module on NYUv2 video data,and achieved dense semantic mapping.