Research On Several Problems In SLAM For Security Robots

Due to the potential safety hazards in the concentration of laboratories such as laboratories in universities and colleges,the intelligent security system of colleges and universities led by technology prevention has gradually developed,and the security patrol robots are the core driving force.In this context,this paper conducts in-depth research on several issues in the core SLAM(simultaneous positioning and map construction)technology for security robots.The major work is as follows:The essential knowledge relevant to visual SLAM is introduced,including pinhole camera model and the relation of camera pose transformation.The depth measurement principle of depth camera is illustrated in detail,which lays a theoretical foundation for the next step of the research work.A depth image enhancement algorithm based on RGB image fusion is proposed to repair the black hole region.Firstly,coordinate mapping is carried out to calculate the corresponding relationship between the color image and the depth image.Secondly,color image and the depth image are preprocessed by morphological method and the effective supporting edges of large void areas are extracted.Thirdly,the invalid areas in depth images are repaired by combining spatial domain,color and just noticeable blur edges.Finally,an improved weighted guided filter is used to denoise minor isolated areas.The results represent that the proposed algorithm has a significant effect on the repair of images with missing depth information,and the sharpness of the edges of the images is higher.A visual odometry based on point-line feature synthesis is proposed to alleviate the feature dependence of scene.Firstly,the ORB feature is selected as the point feature.Subsequently,the LSD segment detection method and LBD segment description method are analyzed and improved.Eventually,the visual odometry frame of the system is built,the error model is defined and the expression of camera pose is deduced.The robustness of the system is verified by experiments.The positioning error in the actual scene is about 1.2%,which meets the real-time requirements.