中文核心期刊
CSCD来源期刊
中国科技核心期刊
RCCSE中国核心学术期刊
重庆交通大学学报（自然科学版）

重庆交通大学学报（自然科学版） ›› 2022, Vol. 41 ›› Issue (12): 11-17.DOI: 10.3969/j.issn.1674-0696.2022.12.02

• 交通+大数据人工智能 • 上一篇    下一篇

基于改进CenterNet的交通场景目标检测技术研究

赵奉奎，成海飞，苏珊珊，张涌   

  1. (南京林业大学 汽车与交通工程学院，江苏 南京210037)
  • 收稿日期:2021-08-12 修回日期:2022-10-16 发布日期:2023-01-16
  • 作者简介:赵奉奎（1986—），男，山东济宁人，博士，主要从事智能车辆环境感知，计算机视觉等方面的研究。E-mail:zfk@njfu.edu.cn 通信作者：张涌（1971—），男，江苏扬州人，博士，主要从事节能与新能源汽车控制，自动变速器控制，智能车底盘线控等方面的研究。E-mail:zy.js@163.com
  • 基金资助:
    江苏省产业前瞻与关键核心技术项目（BE2022053-2）；江苏省重点研发计划（现代农业）（BE2021339）；南京林业大学青年科技创新基金项目（CX2019018）

Traffic Scene Target Detection Technology Based on Improved CenterNet

ZHAO Fengkui, CHENG Haifei, SU Shanshan, ZHANG Yong   

  1. (College of Automobile and Traffic Engineering, Nanjing Forestry University, Nanjing 210037, Jiangsu, China)
  • Received:2021-08-12 Revised:2022-10-16 Published:2023-01-16

PDF

摘要： 针对交通场景目标复杂、遮挡严重，导致目标难以检测的问题，提出了基于改进CenterNet的目标检测算法。CenterNet使用的激活函数ReLU函数在输入值为负值时输出及梯度均为0，导致CenterNet部分神经元无法计算输入信息，LeakyReLU函数在输入值小于0时存在一个很小的值，保留神经元的梯度值，能够解决ReLU函数中神经元死亡的问题。将CenterNet网络中的ReLU激活函数改进为LeakyReLU函数。同时改进了主干网络，综合考虑浮点运算数量和网络深度等因素，将原CenterNet网络主干网络改为ResNet-50，并且增加了空间金字塔池化结构（SPP），能更好提取特征信息，针对交通场景重新训练神经网络，利用测试集数据对神经网络进行测试。试验结果表明：对CenterNet算法进行优化后目标检测的召回率和精确率得到了提高，其综合评价指标F1值平均提高了0.14，mAP由83.18%提高至87.35%，FPS指标虽略有下降，但其他性能指标提升明显，说明改进的CenterNet目标检测模型具有较高的交通场景目标检测能力。

关键词: 交通运输工程；CenterNet；目标检测；神经网络

Abstract: Aiming at the problem of complex targets and serious occlusions in traffic scenes, which made objects difficult to detect, an objects detection algorithm based on improved CenterNet was proposed. The activation function ReLU function used by CenterNet had an output and gradient of 0 when the input value was negative, which caused that some neurons in CenterNet could not calculate the input information. LeakyReLU function had a very small value when the input value was less than 0. The problem of neuron death in the ReLU function can be solved by retaining the gradient value of neurons. The ReLU activation function in the CenterNet network was improved to the LeakyReLU function. Meanwhile, the backbone network had also been improved. Comprehensively considering the factors such as the number of floating-point operations and network depth, the original CenterNet backbone network was changed to ResNet-50 and the spatial pyramid pool structure (SPP) was added, which could better extract feature information. The neural network was retrained for traffic scenarios, and the neural network was tested by using test set data. The test results show that the recall rate and accuracy rate of target detection are improved after optimization of CenterNet algorithm. The comprehensive evaluation index F1 value has increased by 0.14 on average, and the mAP has increased from 83.18% to 87.35%. FPS has dropped slightly, but other performance indicators have improved significantly, indicating that the improved CenterNet target detection model has a higher ability of target detection in traffic scenes.

Key words: transportation engineering; CenterNet; target detection; neural network

中图分类号: