Research On Lightweight Semantic SLAM Based On Deep Learning

Simultaneous Localization and Mapping refers to the robot locating itself in an unknown environment only by the observed features and constructing a map of its surroundings.To address the problems of visual odometry in highly dynamic environments and the problems of traditional loopback detection modules,a visual odometry method based on dynamic object removal,a loopback detection method based on capsule networks,and a lightweight semantic SLAM system based on deep learning are proposed.The main research is as follows.Firstly,the research background and significance of lightweight semantic SLAM are outlined,the domestic and international research results of SLAM technology and semantic segmentation technology are summarized in recent years,The main problems and technical difficulties of the semantic SLAM technology combined with deep learning in dynamic environment are analyzed in detail.And the main research objectives and research directions are formulated.At the same time,the research content and organizational arrangement are briefly described,and the theory and principle concepts of the visual SLAM algorithm used are briefly introduced.Secondly,a visual odometry calculation method based on dynamic object removal is designed to address the problems of poor real-time performance and inaccurate localization of visual SLAM systems in highly dynamic environments.Using Mobile Net v3 as the backbone network.A lightweight semantic segmentation network is formed by using multi-scale cavity space pyramid as the decoding network combined with the channel attention mechanism.Combined with the optical flow method to remove dynamic object interference,the designed visual odometry is made to have high real-time performance and accuracy in a highly dynamic environment,and the superior performance of the designed algorithm is verified by experiments.Thirdly,to address the problems of easy loss of location features and poor processing of complex scenes in the traditional loopback detection network,the capsule network is used to replace the convolutional blocks in the descriptor extraction network so that the descriptors generated by the model contain deeper features,thereby improving the loopback detection accuracy of the model,and the excellent anti-interference and high accuracy of the network in complex environments are experimentally verified.Finally,the designed semantic segmentation module and loopback detection module are used to build a lightweight semantic SLAM system.Moreover,the experimental comparison with various improved algorithms in a variety of highly dynamic environments and data is conducted to demonstrate the excellent performance of the designed SLAM system in highly dynamic environments.