Indoor Scene Semantic Map Construction Based On Deep Learning

With the rapid development of artificial intelligence technology,research on mobile robots will continue to develop towards better interactivity and higher levels of intelligence.To adapt to the future research trends of robots,further research is needed on the relevant technologies of the robot’s existing environment perception methods,especially simultaneous localization and mapping(SLAM)in unknown environments.This paper proposes solutions to the deficiencies of existing map construction and semantic perception methods to provide more accurate environmental and higher-level semantic information to robots,aiming to construct semantic maps with more rich and accurate environmental information.The research work of this paper has important application value and significance for robots to complete interactive tasks for specific targets in unknown environments.The accuracy and diversity of information directly affect the robot’s understanding of the environment.Robots exploring unknown environments face some problems:(1)The initial data extracted by the SLAM front end directly affects the performance of the overall system’s pose estimation.Poor feature information and erroneous feature matching results can lead to estimation errors in the entire system.(2)The hardware computing resources that mobile robots can carry are limited and cannot support the use of deep learning networks with excessively large model parameters to complete image semantic extraction tasks.(3)The map constructed by traditional SLAM has poor interactivity and can only be used for simple robot navigation and obstacle avoidance,which cannot support robots to complete higher-level interactive tasks.To solve the above problems,this paper conducts in-depth research on the semantic map construction method of robots in unknown environments using visual SLAM algorithm theory,deep learning network,semantic information mapping and fusion methods,and some algorithm optimization solutions.The main research contents are as follows:1.Based on the theoretical framework of visual SLAM,using the ORB-SLAM2 algorithm as the basis,this paper adopts a pre-screening feature point detection method and a feature mismatch information elimination method based on the PROSAC algorithm,effectively optimizing the input measurement data of the algorithm and reducing the estimation error of the overall mapping system.A thread for constructing a dense point cloud map is added to the original algorithm,and a method for constructing a three-dimensional point cloud map with higher estimation accuracy and richer map information is proposed.2.Combining the robot scene semantic perception method based on visual information,this paper proposes a semantic extraction method that accurately locates and semantically segments target objects in 2D images through the YOLOv3 target detection network and DeepLabV3+semantic segmentation network.The model parameter size is reduced by using depth separable convolution and lightweight feature extraction network.This method can support robots to accurately and efficiently obtain the semantic information of specific objects during the mapping process.3.Based on the inverse transformation principle of the camera imaging process and the depth information of the scene environment,this paper maps the semantic annotation information to the 3D point cloud map.A multi-view semantic information fusion method based on point cloud centroid distance is proposed.According to the constraint relationship of the SLAM algorithm’s pose estimation,the semantic information of target objects in multiple view keyframes during the mapping process is mapped and fused into the three-dimensional point cloud map.The redundant data with too much duplicate information is eliminated by using the keyframe adaptive grouping algorithm,and the global consistent 3D semantic map construction is finally completed.The methods proposed in this paper have been verified by simulation and experiments,and the results show that the proposed methods are effective and have achieved the expected goals.