欢迎访问《汽车安全与节能学报》,

JASE ›› 2020, Vol. 11 ›› Issue (1): 111-116.DOI: 10.3969/j.issn.1674-8484.2020.01.012

• 汽车安全 • 上一篇    下一篇

智能驾驶实时目标检测的深度卷积神经网络

申恩恩 1,胡玉梅 1,陈 光 1,罗  攀 1,朱 浩 2 #br#   

  1. (1. 重庆大学 机械传动国家重点实验室,400044 重庆,中国; 
       2.四川大学 力学科学与工程系,610065 成都,中国)
  • 收稿日期:2019-04-20 出版日期:2020-03-31 发布日期:2020-04-01
  • 通讯作者: 胡玉梅,教授,E-mail:cdrhym@163.com。
  • 作者简介:第一作者 / First author: 申恩恩( 1995—),男(汉),河南,硕士研究生。E-mail:enenshen@163.com。
  • 基金资助:
    国家自然科学基金资助项目( 51805339)。

Deep convolution neural network for real-time object detection of intelligent-driving

SHEN Enen 1, HU Yumei 1, CHEN Guang 1, LUO Pan 1, ZHU Hao 2   

  1. (1. State Key Laboratory of Mechanical Transmission, Chongqing University, 400044 Chongqing, China; 
     2. Department of Mechanics Science and Engineering, Sichuan University, 610000 Chengdu, China)
  • Received:2019-04-20 Online:2020-03-31 Published:2020-04-01

摘要: 为提高深度学习神经网络运行速度,满足智能驾驶对算法实时性的要求,基于一种一体化实 时目标检测算法 YOLO 和一种目标检测网络模型 Faster RCNN,提出一种结合两者特点的实时目标 检测神经网络。该网络保留区域卷积神经网络(R-CNN)算法的二次检测模式和区域生成神经网络 RPN,去掉先验框,采用YOLO直接预测位置。结合 Mask R-CNN 中的 ROI-Align 方法进行二次 位置修正,减少了 Faster R-CNN 中ROI-pooling 所带来的位置预测偏差。对改进后的网络在 KITTI 数据集上进行测试,结果表明:改进后的神经网络检测一次仅耗时 38 ms,检测的平均精确度高于 YOLO 和Faster RCNN,且对于不同大小的目标都具有很好的泛化能力。

关键词: 智能驾驶汽车 , 环境感知 , 目标检测 , 深度学习 , 视觉感知 , 卷积神经网络(CNN)

Abstract:  A real-time deep neural network for object detection was proposed based on a real object detection model of YOLO and an algorithm of Faster R-CNN to improve the running speed of deep learning neural network and to meet the real-time requirements of the algorithm for intelligent-driving. The neural network retained the secondary detection mode of the R-CNN series and region proposal network (RPN), removed the priori box, and used the YOLO to predict the location directly. The position prediction error caused by ROI-pooling in the Faster R-CNN was decreased, combined with a ROI-Align method in a Mask R-CNN. The improved network was tested on KITTI dataset. The results show that the improved neural network detection takes only 38 ms at once detection, the detection average accuracy of improved networks is higher than YOLO and Faster RCNN with a good generalization ability for objects with different sizes at a faster speed with a higher detection precision.

Key words: intelligent-driving cars ,  , environmental perception ,  object detection , deep learning ,  visual perception , convolutional neural network(CNN)