摘要
车载目标检测系统作为自动驾驶中的重要组成部分,可有效降低交通事故的发生。以卷积神经网络为代表的目标检测算法相较浅层机器学习算法在检测精度方面有了明显提升,但也为低功耗、小型化的嵌入式实现带来了困难。文中使用多核异架构ZYNQ-SOC平台,采用软硬件协同设计的原则,设计了一款车载目标检测系统。针对卷积神经网络结构复杂的问题,提出了一种软硬件划分的改进方法;针对运算量巨大的问题,提出了使用FPGA算法实现卷积运算的并行加速。试验结果表明,该系统设计在各种复杂路况和光照条件下检测结果准确,在功能上体现了良好的鲁棒性。在加速性能方面,相对于ARM cortex-A9吞吐量提高了百倍级,运算效能也达到了60倍左右。在检测精度方面,使用YOLOv2算法在不同检测场景下准确率和覆盖率均超过80%。系统的各项指标达均到了项目设计要求,满足了车载目标检测的应用需求。
As an important part of autonomous driving,the on-board target detection system can effectively reduce the occurrence of traffic accidents.Target detection algorithms represented by convolutional neural networks have significantly improved detection accuracy compared to shallow machine learning algorithms,but have also brought difficulties to low-power,miniaturized embedded implementations.This paper uses the multi-core heterogeneous architecture ZYNQ-SOC platform,and adopts the principle of software and hardware co-design to design an in-vehicle target detection system.Aiming at the complex structure of the convolutional neural network,an improved method of hardware and software partitioning is proposed.For the huge amount of problems,the parallel acceleration of convolution operations using FPGA algorithms is proposed.The experimental results show that the system is designed to accurately detect the results under various complicated road conditions and lighting conditions,and it shows good robustness in function.In terms of acceleration performance,compared to ARM cortex-A9,the throughput has been improved by a hundred times,and the computing efficiency has also reached about 60 times.In terms of detection accuracy,the accuracy and coverage of YOLOv2 algorithm in different detection scenarios are over 80%.All the indicators of the system have met the project design requirements and met the application requirements of vehicle target detection.
作者
李向阳
LI Xiang-yang(School of electrical engineering and automation,TianGong University,Tianjin 300387)
出处
《机械设计》
CSCD
北大核心
2020年第S01期35-38,共4页
Journal of Machine Design
关键词
卷积神经网络
可编程逻辑
软硬件协同设计
硬件加速
convolutional neural network
programmable logic
software and hardware co-design
hardware acceleration
