摘要
考虑到现有的基于检测的多目标跟踪算法多会出现因目标漏检或数据关联算法冗余而造成的目标ID频繁切换、跟踪轨迹断开等问题,提出了无人车驾驶场景下的多目标车辆与行人跟踪算法.首先,选取CenterNet网络作为目标检测器,并用嵌入了1×1卷积和SE-Net的Res2Net来替代网络原有的残差单元,以提升网络对空间信息和通道信息的提取能力,提高目标检测器性能.接着,用孪生网络来提取目标所在区域的特征,进行关联概率度量,再用匈牙利算法对相邻帧目标进行关联.最后,用区域推荐网络设计的辅助跟踪器对漏检或消失又出现的目标进行持续跟踪,并将可靠的跟踪结果合并到轨迹中.实验结果表明,与已有的方法对比,所提方法在KITTI跟踪基准数据集上对于车辆与行人的跟踪具有竞争力.
Considering that the existing multi-object tracking algorithms based on tracking-by-detection framework,they often have the problems of frequent switching of object’s ID and disconnection of tracking track caused by missing detection of object or redundancy of data association algorithm.Thus,this paper proposes a multi-object vehicle and pedestrian tracking algorithm in driving scene of unmanned vehicle.Firstly,CenterNet network is selected as the object detector,and res2 net embedded with 1×1 convolution and SE-Net is used to replace the original residual unit in the network,so as to improve the network’s ability to extract spatial information and channel’s information and improve the performance of the object detector.Then,siamese network is used to extract the features of the region where the target is located,and the probability of association is measured.Then,the Hungarian algorithm is used to match the detected object of adjacent frame.Finally,the auxiliary tracker designed by region proposal network is used to track the missing or disappearing objects continuously,and the reliable tracking results are incorporated into the trajectory.Compared with the existing methods,the experimental results show that the proposed method is competitive for vehicle and pedestrian tracking on the KITTI tracking benchmark dataset.
作者
顾立鹏
孙韶媛
李想
刘训华
宋奇奇
GU Li-peng;SUN Shao-yuan;LI Xiang;LIU Xun-hua;SONG Qi-qi(College of Information Science and Technology,Donghua University,Shanghai 201620,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2021年第3期542-549,共8页
Journal of Chinese Computer Systems
基金
上海市科委应用基础研究项目(15JC1400600)资助。
关键词
机器视觉
目标检测
孪生网络
区域推荐网络
多目标跟踪
machine vision
object detection
siamese network
region proposal network
multiple object tracking
