Research On Visual SLAM Algorithm For Mobile Robots Integrating Semantic Information Of Objects

In recent years,with the progress of society and the development of science and technology,mobile robots have played a more important role in the rich production and life of human beings.The emergence of artificial intelligence has made the functional requirements for mobile robots more diverse.As one of the necessary functions of mobile robots,autonomous positioning and map construction have attracted extensive attention of many scholars.Therefore,in this thesis,aiming at the problem of poor accuracy of simultaneous localization and mapping(SLAM)of mobile robots and single map information,the semantic information of objects in the environment is extracted and integrated into visual SLAM,and the movement of fused object semantic information is carried out.Research on SLAM algorithm for robot vision.Firstly,according to the hardware conditions and the engineering requirements to be met,different semantic information fusion methods are compared,and the optimal scheme is selected,and the introduction of the research content and the overall scheme design of this thesis are completed.Secondly,a visual SLAM algorithm using RGB-D camera as the sensor to improve key frame selection is designed to realize the positioning of the mobile robot itself,and complete the pose optimization and map construction.Based on ORB-SLAM2,the algorithm extracts features from RGB images,performs feature matching on two adjacent frames of images,and performs pose estimation through Pnp algorithm to obtain the current pose of the camera.Appropriate keyframes are selected through the improved keyframe selection algorithm,the back-end optimization is carried out by the beam adjustment method and the Newton-Gaussian method,and the loop closure detection based on the bag-of-words model is used to eliminate the accumulated error of the algorithm.Using the TUM public data set to conduct experiments,the experimental results show that the improved algorithm selects more key frames,has higher positioning accuracy,can complete pose estimation and accurately build dense point cloud maps.Then,an improved YOLOv4 target detection and recognition algorithm based on depthwise separable convolution and inter-channel attention mechanism is designed to detect2 D image semantic information.The algorithm uses CSPDarknet as the basic network for feature extraction,constructs a feature pyramid through Spatial Pyramid Pooling(SPP)and Path Augmentation(PANet),and inputs the obtained features into YOLO Head for target detection and recognition.On this basis,an inter-channel attention mechanism is added,and a depthwise separable convolution is used to replace part of the traditional convolution.While improving the effect of target detection and recognition,the network is more lightweight and suitable for mobile robots with CPU as the control core.Using the Pascal VOC dataset to conduct experiments,the experimental results show that the improved algorithm has higher detection accuracy,smaller model parameters and faster detection speed,meeting the requirements of semantic information extraction.Next,an RGB image segmentation algorithm and a semantic information fusion algorithm are designed to segment and extract the semantic information of objects in two-dimensional images,and integrate them into visual SLAM to construct a deep semantic map.Image segmentation mainly uses the combination of Graphcuts and Grabcut algorithms to achieve accurate segmentation of RGB images.Through the information consistency and visual SLAM positioning results,the pixel inverse operation is performed to construct a deep octree map,and the target object is semantically labeled and updated in real time,and the semantic information of the object is integrated into the visual SLAM to construct a deep semantic map.Using the TUM public data set to conduct experiments,the results show that the designed algorithm can integrate the semantic information of objects into the visual SLAM algorithm,and build a clear and clear deep semantic map.Finally,a mobile robot prototype is designed and built,and the visual SLAM algorithm,target detection and recognition algorithm,and visual SLAM algorithm fusion of object semantic information are carried out through the built mobile robot.The experimental results show that the improved visual SLAM algorithm in this thesis has higher positioning accuracy during the operation of the mobile robot and can realize the construction of dense point cloud map in the laboratory;the improved target detection and recognition algorithm has higher recognition accuracy and better accuracy.Lightweight network model and faster detection speed;the designed semantic fusion algorithm can successfully integrate the semantic information of laboratory objects into visual SLAM to build a deep semantic map.