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） ›› 2021, Vol. 40 ›› Issue (12): 19-26.DOI: 10.3969/j.issn.1674-0696.2021.12.04

• 交通+大数据人工智能 • 上一篇    

道路复杂交通场景下的改进MDnet目标跟踪算法

王小平1，施新岚2   

  1. (1. 重庆城市管理职业学院 大数据与信息产业学院，重庆 401331； 2. 重庆邮电大学 通信与信息工程学院，重庆 400065)
  • 收稿日期:2020-05-25 修回日期:2021-06-29 发布日期:2021-12-27
  • 作者简介:王小平（1973—），男，四川阆中人，教授，主要从事物联网及信息处理方面的研究。E-mail：468077776@qq.com 通信作者：施新岚（1996—），女，广东惠州人，硕士研究生，主要从事图像处理方面的研究。E-mail:shi.xinlan@foxmail.com
  • 基金资助:
    教育部科技发展中心高校产学研创新基金项目（2018A02034）;重庆市教委科技项目（KJQN201803306）

Improved MDnet Target Tracking Algorithm in Complex Traffic Scene

WANG Xiaoping1, SHI Xinlan2   

  1. (1. School of Big Data and Information Industry, Chongqing City Management College, Chongqing 401331, China; 2. School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China)
  • Received:2020-05-25 Revised:2021-06-29 Published:2021-12-27

PDF

摘要： 针对道路复杂交通场景下机动车、非机动车以及行人目标遮挡、目标旋转变换、尺度变换、目标快速运动等因素造成运动目标图像区域模糊，视频图像序列中的光照突变、背景杂波等干扰因素造成运动目标图像特征信息缺失或被破坏等问题导致跟踪目标准确率不足的问题，研究并分析了图像序列中运动目标的时间相关性，改进了MDnet算法。提出一种收集前序图像中目标的光流特性时间信息，利用光流场中运动目标的光流变化信息确定运动目标探索空间，从而预测当前帧目标所在位置的运动目标跟踪算法。在获取运动目标探索空间后，利用小型卷积核以小视野的方式提取探索空间纹理信息，模型训练中采用了边长为(1+70%)r的探索空间，并利用long-term与short-term互补更新方式，每5帧主动更新，出现20个候选框评分低于准确率阈值0.65时被动更新。研究结果表明：在典型的复杂交通道路目标遮挡、目标跟踪干扰与运动模糊场景中，改进后的MDnet运动目标跟踪准确率为95.93%，相比改进前的MDnet，目标跟踪准确率提升了2.01%，并且模型拟合速度、目标跟踪准确率性能表现更优，利用轨迹预测、小型卷积核的改进MDnet能有效提升道路复杂交通场景下的目标遮挡、目标快速运动跟踪性能。

关键词: 交通工程；违章目标跟踪；智能化道路管控；时间相关性；MDnet； 轨迹预测

Abstract: In the complex road traffic scene, the moving target image area was blurred by the factors such as motor vehicle, non-motor vehicle and pedestrian target occlusion, target rotation transformation, scale transformation and target rapid movement; the loss or destruction of the moving target image feature information was caused by illumination mutation, background clutter and other interference factors in the video image sequence, which resulted in insufficient tracking accuracy. In view of the above problems, the temporal correlation of moving targets in the image sequence was researched and analyzed, and the MDnet algorithm was improved. Meanwhile, a moving target tracking algorithm was proposed, which collected the optical flow characteristic time information of the target in the pre-image and used the optical flow change information of the moving target in the optical flow field to determine the exploration space of the moving target, so as to predict the position of the target in the current frame. After acquiring the exploration space of the moving target, a small convolution kernel was used to extract the texture information of the exploration space in a small field of view. In the model training, an exploration space with a side length of (1+70%) r was used, and the long-term and short-term complementary update methods were used. Every 5 frames were updated actively, and 20 candidate frames were updated passively when the score was lower than the accuracy threshold of 0.65. Research results show that in typical target occlusion, target tracking interference and motion blur scenes in complex traffic roads, the moving target tracking accuracy of the improved MDnet is 95.93%, which is 2.01% higher than that of the MDnet, and the performance of model fitting speed and target tracking accuracy is better. The improved MDnet that uses trajectory prediction and of small convolution kernel can effectively improve the performance of target occlusion and target fast motion tracking in complex road traffic scenes.

Key words: traffic engineering; illegal target tracking; intelligent road control; temporal correlation; MDnet; trajectory prediction

中图分类号: