Deep Learning Based Indoor 3D Semantic SLAM

Autonomous mobile platforms,also known as unmanned platforms,are a form of ground mobile robot.Simultaneous Localization and Mapping(SLAM)can enable unmanned platforms such as mobile robots to locate themselves and map the surrounding environment.Combining semantic information with SLAM can not only improve the positioning and mapping accuracy,but also enable it to recognize and understand the surrounding environment.Aiming at the actual needs and new challenges,this paper proposes a method to build a three-dimensional semantic map by combining a fully convolutional neural network optimized with a convolutional neural network and a three-dimensional visual SLAM.This paper completes the task of combining image semantic segmentation and SLAM technology to build a three-dimensional semantic map of the environment.First,RGB-D information will be obtained by using the depth camera Kinect2.Second,three-dimensional visual SLAM will be completed through the Elastic Fusion algorithm.Finally,a three-dimensional map will be constructed.At the same time,a three-dimensional semantic map can be finished by employing the full convolutional neural network(FCN)to implement semantic segmentation.As for the FCN part,a full convolutional neural network combining different feature layer structures by jumping layers is proposed.This type of full convolutional network learning combines rough,high-level information with clear,low-level information.It only displays pooling layers and the prediction layers and omits the intermediate convolutional layers.It is able to retain rough,high-level semantic information and clear,low-level semantic information.At the same time,it improves the accuracy of semantic segmentation.For the combination of FCN and Elastic Fusion,due to the limitation of the number of channels inherent in FCN,it is not possible to directly perform semantic segmentation on maps built by Elastic Fusion.In this paper,FCN and Elastic Fusion can be well combined by adding channels and initializing the first three channels and the fourth channel respectively and then changing the resolution of the depth map image.Experimental results show that for indoor scenes,the algorithm in this paper can get better semantic maps than the current mainstream semantic segmentation Eigen algorithm.