Construction And Research Of Visual SLAM System In Indoor Complex Scenes

With the rapid development of computer vision and artificial intelligence technology,intelligent mobile robots are gradually applied to all walks of life in society.Simultaneous localization and mapping(SLAM)technology based on vision is the core technology for mobile robots to complete intelligence,has been widely concerned by researchers in academia and industry.At present,the visual SLAM system is mainly based on the research of static simple environment.In the indoor complex environment robot application,due to the influence of multiple moving targets,lighting and other factors,the system performance is seriously interfered,resulting in the robot's positioning accuracy and composition effect not meeting the requirements.In addition,most SLAM systems do not have semantic information,which makes the constructed map unintuitive and cannot be well applied to interactive systems and navigation systems.In response to these problems,this article will design the visual SLAM system from the following points.(1)Design and implementation of the positioning module of the visual SLAM system based on dynamic object detection: The visual SLAM system lacks dynamic information processing problems,resulting in poor estimation of the system's pose and map construction accuracy in complex environments.This paper analyzes the dynamic information detection algorithm based on optical flow in detail,and introduces it into the visual SLAM system to solve the positioning of complex environments.Experiments show that the visual SLAM system positioning algorithm based on dynamic information detection can significantly improve the positioning accuracy of the SLAM system.(2)Design and implementation of the target detection module of the visual SLAM system based on the Feature Fusion Detection Network(FFDNet): the map constructed by the general visual SLAM system contains only geometric entities(points,planes,surfaces,etc.),lacking semantics information.Using FFDNet to introduce the target detection module,the SLAM system has semantic analysis capabilities.The target detection algorithm of FFDNet is based on the improvement of SSD target detection algorithm,which can improve the target detection ability without losing the detection speed.(3)Design of the semantic map module of the visual SLAM system in a complex indoor environment: on the basis of the point cloud map,the construction of the Octomap map is completed.Using Octomap and FFDNet target detection algorithms,combined with the semantic database,to achieve semantic information acquisition in a complex indoor environment,construct a semantic map of the SLAM system to better suit robot navigation and advanced semantic tasks.(4)Design and implementation of the indoor complex environment visual SLAM system: the entire SLAM system is built on the ORB-SLAM2(Oriented FAST and Rotated BRIEF SLAM2)system,which has the capabilities of positioning,target detection and semantic map construction.Through standard data sets and actual scenarios,comparative experiments with ORB-SLAM2 system were conducted in static and dynamic environments.The results show that the SLAM system we designed has improved the ORB-SLAM2 in terms of absolute trajectory error and root mean square error,and the system has the ability to obtain semantic information.