基于YOLOv3与卡尔曼滤波的多目标跟踪算法  

MULTI-TARGET TRACKING ALGORITHM BASED ON YOLOV3 AND KALMAN FILTER

作　　者：任珈民 宫宁生 韩镇阳 Ren Jiamin;Gong Ningsheng;Han Zhenyang(College of Computer Science and Technology,Nanjing Tech University,Nanjing 211816,Jiangsu,China)

机构地区：[1]南京工业大学计算机科学与技术学院,江苏南京211816

出　　处：《计算机应用与软件》2020年第5期169-176,共8页Computer Applications and Software

基　　金：国家重点基础研究发展计划项目(2005CB321901);软件开发环境国家重点实验室开放课题(BUAA-SKLSDE-09KF-03)。

摘　　要：为了更全面地解决行人多目标跟踪中现存的问题,使用YOLOv3检测当前帧中的待跟踪目标,利用卡尔曼滤波器根据当前目标的位置预测其下一位置及边界框大小。采用改进匈牙利算法,根据检测边框和预测边框的交并比和颜色直方图进行数据关联与匹配,通过系统不断迭代获得目标的运动轨迹完成跟踪。对于被遮挡的目标,引入基于区域的质量评估网络,联合多帧高质量检测图像,恢复被遮挡部分,提高跟踪准确率。采用2D MOT 2015数据集进行实验,该算法平均跟踪准确度达到了34.4%,较其他算法有明显提升。In order to address the existing problem of multi-target tracking of pedestrian,this paper considers using YOLOv3 to detect the tracked targets in the current frame.The Kalman filter was used to predict its next position and size of the bounding box according to the current position of the target.We used the improved Hungary algorithm to correlate and match the data according to the intersection ratio of detected and predicted borders and color histogram.The target s motion trajectory was obtained through continuous iteration of the system,and the tracking was completed.For the occluded targets,the region-based quality estimation network(QREN)was introduced to solve this problem.It could combine consecutive high-quality detecting frames to restore the occluded part and improve accuracy of tracking.We use 2D MOT 2015 data set for the experiment,and the average accuracy tracking reaches 34.4%,which is significantly improved,compared with other algorithms.

关 键 词：目标跟踪 计算机视觉 卡尔曼滤波器 匈牙利算法 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]