Research On Moble Robot Semantic Map Building System

Robot map building and location is an important research direction in the field of robotics.The grid map and the topological map constructed by the traditional mobile robot map building method can meet the task requirements of robots for location,navigation,and path planning.However,the above two maps can only describe the geometric information and topological information of the environment,and do not include the advanced semantic information of the environment which are crucial for robots to fully understanding the environment and performing more advanced manmachine interaction tasks.For example,a grid map can express the geometry of a room,but it does not indicate what is in the room,nor does it describe the properties and functions of the room.The semantic map building technology can solve this problem and enable robots to understand environmental information at the semantic level which is more similar to the way humans understand the environment.In the future,it must have high application value in intelligent homes,medical assistance and other fields..However,long-term practice shows that the mobile robot semantic map building is faced with many challenges.For example,the depth map obtained by depth cameras directly have a large number of invalid points that affect the calculation of object positions.The detection accuracy of environmental objects is not high.The computational complexity of the image-based target detection algorithm is too high to be able to run in real time on a mobile robot.Besides,how to combine semantic information and target detection information to build an available semantic map.In addition,how to build,configure,and deploy the hardware and software environment of the robot to adapt to the functions of the robot sensor device drive,control,navigation and location,target detection,and network communication is also an engineering problem to be solved urgently.This article has conducted research on the above issues.The main contributions are:(1)Developed a lightweight target detection model based on deep convolutional neural network,which achieved high-efficiency target detection while ensuring detection accuracy.The inter-frame optical flow information in the video stream is used,and the motion information is used to guide the propagation algorithm to reduce the missed detection rate of the detection algorithm.(2)For depth images generated by Kinect or other RGB-D depth cameras,there are a large number of invalid points represented by black and black holes.Using CUDA technology,a real-time depth image repair algorithm is developed.Both the function of deep image restoration and the real-time performance of the algorithm are ensured.(3)This paper uses SLAM technology to realize the underlying location,navigation,and map building functions of mobile robots.Based on this,it uses Bayesian inference framework and integrates the metrics and visual identification information of the environment to build semantic maps..This paper presents an effective solution to the depth image restoration,the realtime of object detection algorithms based on deep convolutional neural networks,and the building and representation of mobile robot semantic maps.Our method is verified by the results of the semantic map building in real indoor scenes.