Research On Key Technologies Of Semantic SLAM Based On SoC

SSemantic slam refers to that on the basis of real-time location and map building,it can recognize the surrounding environment through various methods,obtain the geometric information and semantic information in the environment at the same time,and use these information to create semantic map,improve the intelligent navigation ability of robot.With the development of computer vision in recent years,machines can extract more and more information from images.Therefore,it has become a mainstream method to use visual sensors as information sources.However,the processing of images requires a long calculation time and a large amount of computing resources,so it is generally difficult to apply the processing of images to embedded platforms.In this thesis,the semantic information is obtained by the result of target detection.By integrating visual sensor,IMU and odometer information,more accurate geometric information is obtained.The methods of acquiring geometric information and semantic information are extensively studied,and the hardware architecture based on SoC is designed.The algorithm is accelerated to reduce power consumption.The main contents and innovations of this thesis are as follows:(1)In terms of target detection,this thesis adjusts and optimizes the algorithm based on MobileNet-SSD.This thesis describes in detail the method of simplifying the normalization operation in MobileNet-SSD,and proposes a scheme to convert floating point numbers into fixed ones.As for the characteristics of a large number of convolution operations in the algorithm,this thesis proposes a processing unit for calculating convolution,and has achieved good results.Concerning the three convolutions in MobileNet-SSD,the processing unit has three modes respectively:deep convolution mode,point-by-point convolution mode,and standard convolution one.On the Xilinx Zynq-7035 platform,the experimental verification and analysis of the design scheme make full use of the LUTs in the platform.It can be seen from the experiment that when the system clock frequency is 150 MHz,the target detection algorithm designed in this thesis runs at 13 frames/second,and the power consumption of the whole system is only 5W.It has a high performance ratio and can be used real-time target detection in embedded platforms.(2)In terms of SLAM,this thesis,based on the VINS-mono,further integrates the odometer information by the algorithm and adjusts and optimizes the algorithm structure.By doing so,the mobile robot can obtain more accurate positioning effect.In addition,this thesis uses the SoC design architecture to accelerate the proposed algorithm.Considering that the feature detection and tracking module need high real-time performance and that the module needs a lot of computational characteristics of the image,this thesis designs and explains the hardware implementation of the feature detection and feature tracking module.Via the experimental verification and analysis on the Xilinx Zynq-7020 platform,it is proved that the proposed algorithm and design have higher precision and speed,which can meet the real-time positioning needs of small mobile robots.It has certain practical significance and can be applied to sweeping robots,automatic guided transport vehicle and other scenes.