摘要
在自动驾驶系统中,基于视觉的车辆前方多目标检测和轨迹跟踪能够为前方目标的姿态估计、行为分析提供有效信息。针对协同运动信息和核相关滤波跟踪信息的多目标跟踪的不足,使用卷积神经网络YOLOv2检测目标,提出了融合核相关滤波和目标运动信息的多目标轨迹跟踪方法,目的是将运动信息融入到图像特征跟踪容器中,优化运动模型,减少环境噪声造成的目标跟踪丢失、偏离。提出了基于运动信息改进核相关滤波跟踪尺度不变性算法。建立了多目标的检测跟踪容器,提出了结合目标属性、重合度、运动状态、跟踪状态的多目标匹配方法。实验表明,本文算法能够实现一定场景下的多目标的持续实时轨迹跟踪,平均有效跟踪率为92.5%。
In an automated driving system, vision-based multi-target detection and trajectory tracking in front of the vehicle can provide effective information for pose estimation and behavior analysis of the front target. For the deficiency of multi-target trajectory tracking of integrated motion information and kernelized correlation filter tracking information, the convolutional neural network YOLOv2 is used to detect the target and a multi-target tracking method combining kernel correlation filtering and target motion information is proposed. The purpose is to integrate the motion information into the image feature tracking container so as to optimize the motion model and reduce the target tracking loss and deviation caused by the environment noise. An improved kernelized-correlation filter tracking scale invariance algorithm based on motion information is proposed. A multi-target detection tracking container is established, and a multi-target matching method combining target attribute, coincidence degree, motion state and tracking state is proposed. Experiments show that the proposed algorithm can achieve continuous real-time trajectory tracking of multiple targets in a certain scenario, and the average effective tracking rate is 92.5%.
作者
廖家才
曹立波
夏家豪
张晓
吴强
Liao Jiacai;Cao Libo;Xia Jiahao;Zhang Xiao;Wu Qiang(Hunan University, State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Changsha 410082)
出处
《汽车工程》
EI
CSCD
北大核心
2019年第10期1179-1188,共10页
Automotive Engineering
基金
长沙市科技重大专项支持项目(kq1703024)
大学生研究性学习和创新性实验计划项目(湘教通[2018]255号)资助
