多特征融合的端到端链式行人多目标跟踪网络  

Chained End-to-End Pedestrian Multi-Object Tracking Network with Multi-Feature Fusion

作　　者：周海赟[1] 项学智[2] 王馨遥 任文凯 ZHOU Haiyun;XIANG Xuezhi;WANG Xinyao;REN Wenkai(Institute of Public Security,Nanjing Forest Police College,Nanjing 210023,China;School of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]南京森林警察学院治安学院,南京210023 [2]哈尔滨工程大学信息与通信工程学院,哈尔滨150001

出　　处：《计算机工程》2022年第9期305-313,共9页Computer Engineering

基　　金：中央高校基础科研业务费项目(LGY201802,LGZD202102);国家自然科学基金(61401113);黑龙江省科学基金项目(LH2021F011);华为MindSpore学术基金。

摘　　要：目标检测、特征提取与数据关联作为多目标跟踪网络中重要的组件,独立或部分联合地发挥作用,这种组件分离的方法虽取得了良好的跟踪效果,但增加了跟踪网络的复杂性,影响了跟踪速度。为提升行人多目标跟踪速度及维持跟踪精度,提出一种端到端链式行人多目标跟踪网络。将目标检测、特征提取与数据关联集成到一个统一的框架中,将连续2帧图片组成一个节点作为输入,直接回归出节点之间相同目标的成对边界框,利用相邻节点之间公共帧的强相似性,仅使用交并比匹配进行数据关联,以提高跟踪速度。使用多特征融合的双向特征金字塔,并在金字塔网络中引用改进可变形卷积,提高模型对目标形变的适应性。为解决正负样本不平衡及梯度贡献的差异,将focal loss与BalancedL1 Loss组成多任务学习损失函数以促进网络的均衡学习。在MOT17数据集上的实验结果表明,与DeepSORT、TubeTK、CenterTrack等网络相比,该网络可有效实现跟踪速度与精度的平衡,多目标跟踪精度为69.6,跟踪速度保持为21.6 frame/s。Object detection,feature extraction,and data association as important components in multi-target tracking network,work independently or partially jointly. Despite the improved tracking performance,separated components increase the tracking network complexity and decrease the tracking speed. An end-to-end chained network with multifeature fusion is proposed to increase the speed of pedestrian multi-object tracking while maintaining tracking accuracy.The network integrates object detection,feature extraction,and data association into a framework.Two adjacent frames form a node as the input. The network regresses the bounding box pairs of the same target in the node. The common frames across nodes have a strong correlation such that using Intersection over Union(IoU)matching for data association improves the tracking speed. In addition,the multi-feature fusion pyramid is adopted to fully integrate the high-level semantic information and low-level position information. The pyramid adopts deformable convolution v2,which increases adaptability to the deformation of objects.Focal loss and balanced L1 loss form multitask learning loss for promoting the balanced learning to improve the tracking performance,owing to the imbalance in the positive and negative samples and the differences in the gradient contributions.The experimental results for the MOT17 dataset show that compared with DeepSORT,TubeTK,CenterTrack,and other networks,this network can effectively achieve the trade-off between the tracking speed and accuracy. The tracking accuracy Mota value is 69.6,and the tracking speed is maintained at 21.6 frame/s.

关 键 词：多目标跟踪 链式跟踪 多特征融合 特征金字塔 多任务损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]