Research On The Hardware Acceleration Platform For Semantic Visual SLAM Systems

The semantic visual SLAM(Simultanetic Localization and Mapping)is of great significance for intelligent mobile robots to realize natural interaction and specific tasks.At present,the research of semantic visual SLAM mainly focuses on replacing some modules of visual SLAM with deep learning methods at the algorithm level,which lacks the consideration of system hardware implementation and practical application.At the same time,the deep learning method requires GPU(Graphics Processing Unit)for data processing,which may cause problems such as high-power consumption and large volume.So more efficient computing architecture and hardware platform are needed to improve the overall processing efficiency.Considering the portability and high parallelism of FPGA(Field Programmable Gate Array),this paper proposes a design for semantic visual SLAM system based on the HERO platform(Heterogeneous Extensible Robot Open Platform),which takes FPGA as the computing core.In this design,a typical semantic visual SLAM system—DS-SLAM(Semantic Visual SLAM Towards Dynamic Environment)is taken as an example to study the optimization and acceleration of semantic segmentation network,system integration and robot deployment.The main work and innovations are summarized as follows:(1)According to the requirements of indoor application scenarios and FPGA hardware implementation,the Seg Net network of semantic thread is optimized in DSSLAM system.The Seg Net network is trained according to different object categories,and the network model suitable for the system is obtained.Through the combination of Batch Normalization layer and Convolutional layer,the computation of inference is reduced.Through the 8bits fixed-points quantization,the number of network parameters is effectively reduced.GA(Global Accuracy),CA(Class Accuracy)and MIo U(Mean Intersection over Union)are only reduced by 0.817%,1.102% and 1.56%.(2)Based on the existing hardware acceleration architecture of semantic segmentation network,the hardware implementation of Seg Net network is completed in DS-SLAM system.Through the research of parallelization,pipeline structure,kernel execution and data access,the processing speed of the system is improved.The energy efficiency has been improved by 41 times on FPGA development platform compared to the traditional CPU platform.(3)An integration scheme of DS-SLAM system is designed based on the HERO platform,which realizes the deployment of mobile robot(Turtlebot2)and the creation of semantic octree map.Experiments on datasets and real scenarios shows that the proposed scheme ensures the location accuracy of DS-SLAM system and improves the energy efficiency by 50% compared with the traditional computing architecture.Through the research on the hardware acceleration technology for semantic visual SLAM system in this paper,the system gets rid of the dependence on high-power computing platforms.It provides a high energy efficiency solution for localization and mapping of mobile robot.