Feature Extraction Algorithm For Unmanned Laser Point Cloud And Its FPGA Acceleration

Three-dimensional lidar has gradually become an important sensor in the field of unmanned driving due to its high ranging accuracy,unaffected by light,and good real-time performance.It plays an important role in scenes such as high-precision positioning,obstacle recognition and classification.In these application scenarios,feature extraction technology is generally used to extract the three-dimensional features of the scanning environment to achieve the purpose of simplifying the calculation and improving the accuracy.However,the current 3D lidar point cloud feature extraction is mostly based on the CPU or GPU platform,so there are disadvantages such as high power consumption,large volume,and low energy efficiency ratio,which are not conducive to vehicle-level applications.Aiming at the above problems,this paper designs a scheme based on FPGA to realize the acceleration of laser point cloud feature extraction algorithm.Analyzed and compared the advantages and disadvantages of CPU,GPU,FPGA and ASIC in processing operations,and finally adopted the "ARM+FPGA" design method,using the Zynq UltraScale+MPSoC series chips introduced by Xilinx,and integrated programmable logic and processing system on a single chip,through the method of software and hardware co-design to achieve the extraction of feature points of 3D laser point cloud,while using FPGA to bring faster calculation speed,but also through the ARM to achieve the flexibility of the system,Under the premise of ensuring performance,greatly reducing power consumption.The research content of this article mainly includes the following aspects:(1)In-depth study of the working principle of 3D lidar and the basic theory of laser point cloud feature extraction algorithm.On this basis,perform detailed analysis of the laser point cloud filtering,ground point cloud segmentation,non-ground point cloud clustering,corner feature point and surface feature point extraction,the defect removal algorithm,optimize the algorithm implementation process,and lay the foundation for the algorithm’s FPGA implementation.(2)In-depth study of the ARM+FPGA heterogeneous platform,and provide a software and hardware co-design solution implemented by FPGA for the 3D laser point cloud feature extraction algorithm.Use SDSoC development tools to analyze the computation and performance bottlenecks of the algorithm,and make full use of the performance advantages of FPGA to improve the algorithm;Reasonable allocate software and hardware tasks,use FPGA to achieve hardware acceleration for tasks with high repetitiveness and large calculation;Tasks with small quantity and flexible scheduling are realized by ARM processor,and data interaction is completed through the high-performance data bus between ARM and FPGA,and the optimization and acceleration of the algorithm are realized.(3)In-depth study of SDSoC development environment,skilled use of optimization methods in this development environment,greatly shortening the development cycle of algorithm acceleration.Optimize functions such as point cloud preprocessing and curvature calculation,use loop pipeline,loop unrolling,interface optimization,data bit width optimization and other methods to optimize program running time;analyze the performance report provided by SDSoC software,rationally plan the use of hardware resources,and obtain more information Good acceleration effect;at the same time,the hardware debugging method is introduced in detail,and the data throughput rate,bit error rate,instruction delay and other aspects are analyzed.(4)Based on the ZCU104 platform,complete the realization and test verification of the 3D laser point cloud feature extraction algorithm.Use the point cloud data of three real scenes in the office,elevator and square to conduct experimental tests,compare with the results of laser point cloud extraction using CPU and TX2,and conduct a comprehensive analysis of the extraction results,running time,resource consumption and power consumption.The final test results show that the running time of the laser point cloud feature extraction algorithm accelerated by the ZCU104 platform is 13.2ms,the running time of NVIDIA Jetson TX2 is 65.7ms,and the running time of Intel(R)Core i7-10710U is 26.7ms,At the same time,it is measured that the maximum power consumption of the ZCU104 development board is 10.32W,and the maximum power consumption of the chip is 2.1 W,which is much lower than the power consumption of general-purpose processors.It is verified that the Zynq UltraScale+MPSoC platform has the advantages of low power consumption and high efficiency,can meet the requirements of unmanned driving algorithm acceleration,and has a wide range of application scenarios.