Research On Generation And Recognition Of Visual SLAM-based Indoor 3D Image

Visual SLAM refers to the simultaneous localization and mapping from the camera perspective.With the rapid development of artificial intelligence and science and technology,intelligent robot emerges as the times require.It perceives and analyzes problems in an unknown environment,making visual SLAM a popular research direction in the field of computer vision.Cognition based on visual SLAM,as the premise of realizing complex tasks,such as autonomous navigation,obstacle avoidance,human-computer interaction,etc.,has been widely studied and attracted global attention.In recent years,the in-depth learning has made great breakthrough in the field of computer vision.Based on current studies of environment perception and understanding of visual SLAM,the application scheme of in-depth learning in visual SLAM in view of dynamic indoor environment has been explored and improved in this thesis.The major researches and their findings are as follows:1.In the dynamic environment,visual SLAM is more likely to be affected by the moving objects.Aiming to solve the problem that the position of camera cannot be estimated accurately,a visual odometry system based on in-depth learning of silhouette segmentation is put forward.The high recognition rate of object detection in in-depth learning is combined with depth of field to segment the object accurately.By setting the dynamic probability of an object manually,the possibility of relative motion is judged through frame-to-frame comparison,and its probability as a dynamic object is increased.The final value of target probability is a decisive factor to judge whether the object is dynamic or not.If it is,the points that confirm the dynamicity within the region should be removed,so as to enhance the robustness of system position estimation and the quality of the map.2.For the sake of solving the problem that using traditional in-depth learning to detect dynamic objects in front-end visual odometry has increased time consumption,the lightweight network mobilenet-yolov3 is introduced to detect the object.With fewer parameters,the application of mobilenet-yolov3 improves the efficiency to a certain extent,thus shortens the time required by the front-end visual odometer.3.In order to enhance the accuracy of object recognition,the researches on optimal detection algorithm is carried out,and the object detection model that combines in-depth learning-based detection and active contour without edge(ACWE)is proposed.After the bounding box is enlarged,the active contour without edge(ACWE)is used to detect the edge of each object in depth maps,hoping to improve the accuracy of segmentation,and obtain relevant high-order information of objects,such as category,location,etc.Therefore,a solid foundation for SLAM mapping is laid.4.By setting up a visual SLAM system,a globally consistent 3D point cloud is presented,which is further combined with the high-order information obtained from object detection.Finally,the 3D geometric information of indoor environment and the description of high-order information in a certain sense are realized.The results of system test have shown that,the map constructed in this paper includes both 3D geometric information and certain description of high-order information,which is of great significance to promote robot's cognition and understanding of surrounding environment.