Research On Visual SLAM Algorithm Based On Convolutional Neural Network

With the deepening and development of intelligent concepts in various fields of modern society,driverless and intelligent industrial robots have become one of the current research focal points.One of the more widely used techniques is vision-based Simultaneous Localization and Mapping(SLAM).As most current visual SLAM algorithms are assumed to work in ideal static environments,which rarely exist in real life,it is especially important to make visual SLAM algorithms effectively determine their own positions and perceive the environment in real dynamic environments.Firstly,a semantic segmentation network model based on convolutional neural network is designed in this paper.Considering that most of the current semantic segmentation network models take a longer time to process each frame and have poor real-time performance when embedded into visual SLAM algorithms,FNet reduces the network size by making the encoder part of the network lighter.At the same time,the decoder part of the FNet is introduced to deepen the characteristics of the method to ensure its segmentation accuracy.Secondly,a visual SLAM algorithm based on a lightweight semantic segmentation network(Implement robust dynamic SLAM using a new semantic segmentation,IRD-SLAM)is proposed.The ORB-SLAM2 algorithm is susceptible to the problem of poor positioning accuracy due to interference from dynamic objects in dynamic environments.Therefore,FNet is introduced into the visual SLAM algorithm to segment dynamic objects,while combining the multi-view geometry approach together to eliminate the influence of dynamic objects on the positional estimation.Thirdly,a dense point cloud map construction algorithm based on unsupervised learning method is designed.Considering the constructed dense point cloud map without higher level of understanding,the original dense point cloud map is improved with unsupervised learning method so that the dense point cloud map can have simple semantic information.Finally,the mobile robot and Kinect2 camera are used to build a mobile platform to collect real indoor dynamic environments for experimental tests related to pose estimation and dense map construction.Extensive experimental results show that the IRD-SLAM algorithm designed in this paper enables the mobile robot to operate in a dynamic environment with significantly improved positional estimation accuracy,and at the same time can accurately construct dense point cloud maps with semantic information in a real dynamic environment.