摘 要：针对基于卷积神经网络的可行驶区域检测方法计算耗时长、实时性差等问题，基于 Vitis AI 为其设计了一种定制计算系统，并通过采用模型定点化、网络剪枝、硬件定制等优化方法，实现了对可行驶区域检测方法的高效计算。实验结果表明，在 Xilinx ZCU102 异构计算平台上，可编程逻辑部分的工作频率为 200 MHz 时，所实现的可行使区域检测系统的识别帧率可达到 46 FPS，计算性能可达 903 GOPS，能效比为 50.45 GOPS/W，可以较好地满足实际系统的需求。
关键词：现场可编程门阵列；Vitis AI；可行驶区域检测；定制计算系统；卷积神经网络
DOI:10.19850/j.cnki.2096-4706.2022.01.020
基金项目：国家自然科学基金资助项目（61972180）
中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2022）01-0073-06
Design of Customized Computing System for Drivable Area Detection Based on Vitis AI
LI Huilin1, CHAI Zhilei 1,2
(1.School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China; 2.Jiangsu Provincial Engineering Laboratory of Pattern Recognition and Computational Intelligence, Wuxi 214122, China)
Abstract: Aiming at the problems of long calculation time and poor real-time performance of the drivable area detection method based on convolutional neural network, a customized computing system is designed based on Vitis AI, and optimization methods such as model fixedpointization, network pruning, and hardware customization are adopted, which realizes the efficient computing of the drivable area detection method. The experimental results show that on the Xilinx ZCU102 heterogeneous computing platform, when the operating frequency of the programmable logic part is 200 MHz, the recognition frame rate of the realizable area detection system can reach 46 FPS, the computing performance can reach 903 GOPS, and the energy efficiency ratio is 50.45 GOPS/W, which can better meet the needs of the actual system.
Keywords: field programmable gate array; Vitis AI; drivable area detection; customized computing systems; convolutional neural network
参考文献：
[1] CHEN T，CHEN B D，ZHANG X，et al. Free Space Detection Using Stereo Confidence Metrics and Obstacle Position Probability Maps [C]//2018 14th IEEE International Conference on Signal Processing（ICSP）.Beijing：IEEE，2018：1071-1075.
[2] FAN R，WANG H L，CAI P D，et al. SNE-RoadSeg：Incorporating Surface Normal Information into Semantic Segmentation for Accurate Freespace Detection [C]//Computer Vision–ECCV 2020. Glasgow：springer，2020：340-356.
[3] 李升波，关阳，侯廉，等 . 深度神经网络的关键技术及其在自动驾驶领域的应用 [J]. 汽车安全与节能学报，2019，10（2）：119-145.
[4] CHEN Z，ZHANG J，TAO D C. Progressive LiDAR adaptation for road detection [J].IEEE/CAA Journal of Automatica Sinica，2019，6（3）：693-702.
[5] WANG H L，FAN R，CAI P D，et al. SNE-RoadSeg+：Rethinking Depth-normal Translation and Deep Supervision for Freespace Detection [J/OL].arXiv：2107.14599 [cs.CV].（2021-07-30）.https://arxiv.org/abs/2107.14599.
[6] AHAMAD A，SUN C C，NGUYEN H M，et al. Q-SegNet:Quantized deep convolutional neural network for image segmentation on FPGA [C]//2021 International Symposium on Intelligent Signal Processing and Communication Systems（ISPACS）.Hualien City：IEEE，2021：1-2.
[7] LYU H R，AN F W，ZHAO S R，et al. A703.4 GOPs/W Binary SegNet Processor with Computing-Near-Memory Architecture for Road Detection [J].IEEE Design & Test，2020，39（2）：74-83.
[8] 陈辰，柴志雷，夏珺 . 基于 Zynq7000 FPGA 异构平台的YOLOv2加速器设计与实现 [J].计算机科学与探索，2019，13（10）：1677-1693.
[ 9] SHELHAMER E，LONG J，D A R R E L L T. F u l l y Convolutional Networks for Semantic Segmentation [J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 39（4）：640-651.
[10] ZHANG J L，LI J. Improving the Performance of OpenCLbased FPGA Accelerator for Convolutional Neural Network [C]//The 2017 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays. Monterey：Association for Computing Machinery，2017：25-34.
[11] CHEN Y H，KRISHNA T，EMER J，et al. 14.5 Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks [C]//2016 IEEE International Solid-State Circuits Conference（ISSCC）.San Francisco：IEEE，2016：262-263.
[12] BIANCO S，CADENE R，CELONA L，et al. Benchmark Analysis of Representative Deep Neural Network Architectures [J].IEEE Access，2018，6：64270-64277.
作者简介：李慧琳（1997—），女，汉族，湖南郴州人，硕士研究生在读，研究方向：嵌入式系统；柴志雷（1975—），男，汉族，山西新绛人，教授，博士，研究方向：软件定义的高效计算机系统、嵌入式系统、软硬件协同设计等。