| With the constant improvement of computing capabilities on the heterogeneous embedded platforms,more and more computer vision technologies are applied to embedded devices.Simultaneous localization and mapping(SLAM),based on computer vision technology,has been constantly developed and becomes a hot research topic in recent years.One development direction of SLAM is to combine with the deep learning which can use semantic information to improve the positioning accuracy of SLAM.However,current semantic SLAM usually cannot achieve real-time performance on the heterogeneous embedded platform.Besides,SLAM also has a problem that the positioning of devices is mainly relative to the starting position,so it is not possible to confirm the exact location in real life(that is,global localization).To achieve semantic-based global SLAM on the heterogeneous embedded platform in real-time,this paper designs and implements a front-back-end system for localization in the large-scale road scenes.The main research work is as follows:(1)Research and compare related algorithms.Analyze and compare a variety of depth map algorithms,and finally,select the DispNet which can generate depth maps without holes to obtain depth maps;test and compare the computing speed and accuracy of descriptors among feature point algorithms,and finally,the CudaSIFT which can achieve high-accuracy in real-time is integrated into the system;test and analyze a variety of semantic segmentation algorithms,and finally,choose the DeepLab to implement the semantic function;analyze the performance of traditional SLAM and the localization method with the global map to get the inspiration for the system design in this paper.(2)Design and implement a brand-new front-back-end-based system.The system consists of two parts: the back-end running on a high-performance PC and the front-end running on a heterogeneous embedded platform.The back-end is mainly responsible for offline construction of global maps and providing maps to the front-end through wireless communication;the front-end is mainly responsible for completing real-time localization based on the global map.To solve some problems caused by the splitting of the global map,this paper also proposes a two-level localization method,which can effectively locate the initial position of the front-end to receive the corresponding sub-maps.For the integration of semantic information,this paper proposes a method which uses semantic segmentation to distinguish dynamic objects and static objects in the scene and then remove dynamic objects,to solve the problem that dynamic objects are inconsistent in mapping and localization that they are at two different periods.(3)Optimize and improve some algorithms on embedded devices.It includes the optimization of DeepLab's network structure,the parallel speedup of the 2D-3D matching algorithms,and the use of parallel threads to optimize some computing operations in the localization processing.In addition,this paper considers the heterogeneous characteristics of embedded devices in the selected algorithms and optimization methods,that is,make full use of CPU/GPU resources on embedded devices to achieve better performance.After the system is implemented,it can complete real-time localization on the KITTI dataset at a speed of nearly 11 frames per second,and the final trajectory error of localization is maintained within 1 meter,which takes into account both the real-time performance and localization accuracy. |
PDF Full Text Request