Research On Loop Closure Detection Algorithm Of Indoor Dynamic Scene VSLAM Based On Deep Learning

Simultaneous localization and mapping(SLAM)technology is the key technology to realize autonomous navigation of robots.Aiming at the limitations caused by factors such as scenes where there may be dynamic objects in the room and the inability of traditional maps to understand the environment,a new method based on semantic information is explored.The visual SLAM loopback detection algorithm can improve the object recognition rate,improve the accuracy of the loopback detection under the premise of ensuring real-time requirements,and use the image semantic segmentation technology to integrate the traditional visual SLAM system to establish a semantic map with the ability to understand the environment.It lays a research foundation for the realization of autonomous navigation technology of indoor mobile robots.First,the DeepLabV3+ network is used to obtain the semantic information of the image,and the backbone network is replaced with an improved lightweight feature extraction network,which improves the network training and prediction speed with a small loss of segmentation accuracy.Secondly,in view of the problem that the traditional bag-of-words model only pays attention to the feature point information of the picture,but ignores the type,shape and spatial information of the object,a loop closure detection algorithm based on semantic segmentation is proposed.The spatial and appearance similarity of the image can be obtained by the moment,which can not only improve the accuracy of loop closure detection,but also enhance the mobile robot’s ability to understand the environment.Finally,to filter out the dynamic parts in the indoor scene,a semantic segmentation thread is added to the ORB-SLAM2 framework,combining the real-time semantic segmentation network DeepLabV3+ with motion consistency detection,and then matching the feature points from the detected dynamic regions.Therefore,the robustness and accuracy of the system in dynamic scenes are improved.Based on this,dense semantic octree maps that can be used for robot navigation and complex tasks are created.