Research On Semantic SLAM Based On Deep Learning

Synchronous Localization and Mapping,or SLAM(Simultaneous Localization and Mapping),is a pivotal modus for autonomous robot localization and has received wide attention.Most of the current SLAM algorithms are based on geometric features,which are not highly accurate and robust enough.With the emergence of new technologies such as artificial intelligence,robots are required to be able to perceive not only the geometric information of the environment,but also the ability to perceive more advanced information.And the development of deep learning technology provides new research ideas for the ability of robots to perceive semantic information of the environment.Therefore,the study in this dissertation is focused on visual SLAM and semantic segmentation based on deep learning.The main research contents are demonstrated underneath:(1)A visual SLAM algorithm with a depth camera as the sensor is inspected.The algorithm extracts feature points from adjacent images and matches them,estimates the camera pose using a 3D-2D point-to-point method,and uses a bundle set adjustment algorithm to optimize and solve the camera drift issue by loop discernment eradication.The investigational consequence show that the visual SLAM algorithm with depth camera as the sensor has good stability and can perform feature extraction and matching quickly and faultlessly to satisfy the requirement of real-time robot behavior.(2)A loopback detection algorithm based on VGG16 network is given for the traditional loopback detection algorithm based on manually designed geometric features,and environmental changes such as illumination can affect its accuracy and processing time.The algorithm extracts features based on VGG16,selects the pooling layer at the end of the network as the global feature representation of the image,and then confirms the loop by determining the idiosyncrasy resemblance through the conscious hashing algorithm.The investigational consequence show the reliability of the convolutional neural network for loop detection,and the loop detection accuracy and real-time performance are conspicuously ameliorated.(3)A semantic segmentation algorithm based on Seg Net model is given to meet the demand of robots with semantic information perception capability.The algorithm adopts an encoding and decoding structure,which solves the problems of most semantic segmentation models with complex structure,more parameters and long training time.The coding-decoding structure is used,where the coding network extracts features,the decoding network upsamples and then deconvolutes,and finally the pixel classification image with semantic labels is obtained.The experimental results show that the algorithm achieves 99.16% accuracy for segmentation on the dataset,and achieves good segmentation results to meet the semantic information perception needs of the robot.