Semantic SLAM Algorithm Design And System Implementation Based On Federated Deep Learning

With the rapid innovation of artificial intelligence,robotics and autonomous driving,visual SLAM technology has become the focus of current research.However,the existing visual SLAM system still has many defects.On the one hand,visual SLAM is easily affected by changes of illumination,which affects the accuracy and stability of localization.On the other hand,when there are dynamic objects in the environment,the existing algorithms often fail to track because the algorithm cannot determine whether the camera itself has moved or the external conditions have changed.Considering the efficiency of mapping,in the face of large-scale scenarios,the efficiency of mapping with a single sensor device is low,which has a greater impact on the practicability of the SLAM system.Aiming at the problem that the algorithm is not robust in lighting changes and dynamic environments,this thesis designs a semantic SLAM system based on learning descriptors.This thesis applies deep learning feature extraction method to design a more robust visual odometry under lighting changes.Combining the semantic segmentation information of the image,this thesis designs a feature motion probability update strategy to reject dynamic feature points.As a result,the method minimizes the impact of moving objects on localization and mapping.Finally,a semantic map of point cloud with semantic tags is established based on the semantic information,which can provide support for higher semantic tasks such as robot navigation and environment understanding.Aiming at the problem of low efficiency of single-sensor device mapping,this thesis designs a distributed SLAM system based on the cloud robotic system.The system allows multiple robots to collaborate to build maps.Then it merges the local maps uploaded from each robot on the cloud workstation,which effectively improves the efficiency of map creation.At the same time,this thesis adopts the architecture of federated learning to maintain private models and images on the robot side,and only send bag of words data to the parameter server in the cloud.It effectively reduces the communication bandwidth.For the algorithm and system proposed in this thesis,this thesis carries out experiments to verify the effectiveness of the proposed method.For the first work,this thesis conducts localization test experiments on multiple representative public datasets.The experimental results show that the proposed method is more robust than the traditional method in complex environments.In addition,this thesis also shows the mapping results of the proposed system.For the second work,this thesis first models in the ROS Gazabo environment and map the environment through the proposed algorithm.The experiment proves the feasibility of our proposed method.Besides,this thesis also carries out experiments in the real environment.Through the navigation experiment of the robot car in the built map,the experiment further proves the effectiveness of the map built by the proposed methods.