Research On Dynamic Semantic Multi-graph SLAM Algorithm Based On Monocular Vision

With the development and application of deep learning technology and robotics technology,people need more intelligent robots,and synchronous localization and mapping(SLAM)technology is the basic ability of intelligent robots applied to work and life.In recent years,with the rapid development of SLAM technology,various SLAM algorithms have achieved good performance in some aspects and can be applied to some specific scenarios,but there are still many problems to be solved.Therefore,this topic conducts in-depth research on some existing problems of SLAM algorithm,and proposes a complete set of semantic SLAM algorithm that can perform semantic analysis of scenes in dynamic environment.The algorithm mainly includes the following two parts:(1)Traditional visual SLAM algorithm is under the condition of rigid unchanged what-if scenarios to build figure,result in unable to realize dynamic environment built figure,and could not be overcome environmental characteristic is not obvious or robot was "hijacked" kidnapping scenarios with problems,put forward a more dynamic environment can simultaneous localization and map building(DE-SLAMM)SLAM algorithm.The algorithm firstly introduces a multi-map construction idea.When the algorithm is lost,the algorithm adaptively generates a new local map,and fuses the new maps to form a global map during the loop,so as to solve the problem that the algorithm cannot reconstruct after losing.Secondly,deep learning and multi-view geometry are combined to realize real-time detection of dynamic objects in the environment,and multi-frame fusion technology is used to repair the background of the occlusion part of dynamic objects,so as to solve the problem that traditional SLAM algorithms cannot track and build images in the dynamic environment.Finally,the proposed algorithm is applied to practical scenarios for testing.The results show that compared with the classical V-SLAM(ORBSLAM2,ORBSLAMM and Dyna-SLAM)algorithms,when tracking loss occurs,the proposed algorithm can quickly reconstruct the map and realize the fusion of continuous tracking and new map in a very short time.While ORB-SLAM2 and DYn A-SLAM directly repositioned after the algorithm was lost,and could not continue to track and map the environment they were in.Although ORBSLAMM can continue to build maps after losing ORBSLAMM,its established map cannot realize multi-map fusion and cannot build an overall map.Furthermore,through dynamic environment test experiments,it is found that only the proposed algorithm can realize real-time detection and background repair of all dynamic targets(prior and moving targets),dyn A-SLAM can only realize prior target detection,while the other two algorithms cannot realize target detection and map construction in dynamic environment.(2)However,traditional SLAM algorithms cannot recognize objects in the environment.In order to enable robots to perform semantic analysis of the environment and better recognize objects in the environment,This paper proposes a novel real-time semantic segmentation framework for muti-channel Deep Weighted Aggregation Net(MCDWA Net).With the help of DE-SLAMM algorithm,the robot can accurately segment and recognize objects in the environment during the process of tracking and mapping the environment.In this framework,the idea of multi-channel is firstly introduced to construct a three-channel semantic representation model,which can be used to extract three kinds of complementary semantic information: 1)the low-level semantic channel outputs the local features of the object in the image,such as edge,color and structure;2)Assist semantic channel to extract transition information between low-level semantic and high-level semantic,and realize multi-level feedback to high-level semantic channel to ensure the speed and accuracy of high-level semantic channel extraction;3)The advanced semantic channel obtains the contextual logical relation and category semantic information in the image.Then,a weighted aggregation module of three kinds of semantic features is designed,which can deeply fuse the complementary semantic features of three channels to generate the global semantic description of the image,thus greatly improving the segmentation accuracy of the network.Finally,an enhancement training mechanism is introduced to enhance the characteristic representation of the training stage and to strengthen and improve the training speed.Experimental results show that the proposed method not only has fast inference speed,but also has high segmentation accuracy in complex scenes,and can achieve the balance of semantic segmentation speed and precision.The enhanced training module designed can improve the performance of network training.Finally,a hardware testing platform was designed,and THE DE-SLAMM algorithm and MCDWA_Net algorithm were fused into dynamic semantic multi-graph SLAM algorithm using ROS robot operating system.The whole dynamic semantic SLAM algorithm was tested on the hardware testing platform,so as to verify the running effect of the algorithm on the hardware platform.The hardware test results show that the dynamic semantic multi-graph SLAM algorithm studied in this paper can track the environment and build 3d point cloud map of the environment,and can perform semantic analysis of the environment,with good overall effect when running on the hardware test platform.