Research And Application Of Visual SLAM Algorithm Integrating Target Semantic Information Of Indoor Scenes

With the aging of the social population and the outbreak of the new coronavirus epidemic,the demand for human-computer interactive intelligent mobile robots is increasing.Efficient and accurate semantic maps are the guarantee for the operation of intelligent mobile robots,among which Simultaneous Localization And Mapping(SLAM)as a key technology has become a hot topic in research.However,in the indoor scenario,the map environment features of the traditional SLAM scheme are few,the semantic information of the target in the scene is missing,and the human-computer interaction performance is poor.In order to meet the needs of indoor scene understanding of intelligent mobile robots,this paper proposes an algorithm that can effectively construct indoor semantic maps,and the main research contents are as follows:Firstly,the visual SLAM algorithm in indoor scenes is studiedl.The ORB-SLAM2 algorithm is optimized,aiming at the problems of few features of weak texture areas,great influence by dynamic targets and low performance of loopback model,this paper proposes a feature enhancement strategy for weak texture areas,optimizes the distribution of feature points,introduces dynamic feature detection mechanism to reject dynamic objects,optimizes the closed-loop detection model,and effectively improves the performance of drawing.Experiments on the TUM dataset show that under the combined effect of the three improvements,the accuracy of ATE index is improved by13.19%,and the accuracy of RPE index is improved by 12.57%.Secondly,the indoor target semantic information extraction algorithm is studied.The YOLOv4 algorithm is optimized,in view of the optimization of YOLOv4 activation function and the problem of poor multi-scale expression of indoor targets,through the typical activation function comparison experiment,the Leaky Re LU function is selected as a new activation function scheme,the attention fusion mechanism is introduced,and the classification accuracy of the network is improved.Experiments on the self-made indoor scene dataset of indoor coco show that the detection accuracy reaches 42.09%,which improves the accuracy of semantic informationThen,a visual SLAM algorithm integrating indoor target semantic information is proposed,in order to further eliminate the dynamic target features and make up for the lack of map semantic information,the improved YOLOv4 network is used to extract the target semantics of the indoor images,and the dynamic target feature culling strategy is improved,and the target semantic library is constructed to achieve the semantic map construction.Experiments show that the introduction of semantic information to eliminate dynamic target features can effectively improve the accuracy of drawing,especially in the freiburg3＿walking＿static sequence,the accuracy of ATE evaluation index is improved by 34.07%,and the indoor semantic map is completely constructed by constructing the target semantic library.Finally,the semantic map algorithm of the design is applied to the intelligent mobile terminal robot and runs its performance in the actual environment,and the experiment shows that the semantic map algorithm proposed in this paper can effectively process the motion target in the indoor environment and construct the semantic map of the environment.