基于遮挡感知的行人检测与跟踪算法  

Pedestrian detection and tracking algorithm based on occlusion-aware

作　　者：李颀[1] 王娇[2] 邓耀辉 LI Qi;WANG Jiao;DENG Yaohui(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi’an 710021,China;School of Electrical and Control Engineering,Shaanxi University of Science and Technology,Xi’an 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021 [2]陕西科技大学电气与控制工程学院,陕西西安710021

出　　处：《传感器与微系统》2023年第4期126-130,共5页Transducer and Microsystem Technologies

基　　金：西安市科技计划资助项目(201806117YF05NC13(1));西安市未央区科技计划资助项目(201305);陕西科技大学博士科研启动基金资助项目(BJ13-15)。

摘　　要：针对行人遮挡以及跟踪过程中ID频繁切换的问题,提出一种基于遮挡感知的行人检测与跟踪算法。利用遮挡感知算法提高遮挡情况下的行人检测精度;为提高遮挡情况下的跟踪准确度,融合时变运动特征和梯度方向直方图(HOG)特征为一个新的度量匹配信息,对上下帧的行人多目标进行匹配。实验结果表明:该行人多目标检测算法较基准算法在CrowdHuman测试集上准确率提高了7.49%,召回率提高了9.26%。跟踪算法在MOT数据集上具有较好的鲁棒性和跟踪准确度,相较于近年来流行的YOLOv3-DeepSort,SSD-Sort,MOTA分别提高了5.39%,8.17%。Aiming at the problem that pedestrian occlusion and frequent switching of ID in the process of continuous tracking,a pedestrian detection and tracking algorithm based on occlusion-aware is proposed.The occlusion-aware algorithm is used to improve the precision of pedestrian detection in the case of occlusion;in order to improve the tracking accuracy in the case of occlusion,the time-varying motion features and histogram of oriented gradient(HOG)features are fused as a new metric matching information to match the pedestrian multi-targets in the consecutive frames.The experimental results show that compared with the benchmark algorithm,the accuracy of the pedestrian multi-target detection algorithm in the CrowdHuman test set is improved by 7.49%,and the recall rate is increased by 9.26%.The tracking algorithm has good robustness and tracking accuracy on MOT datasets,which is 5.39%and 8.17%higher than that of YOLOv3-DeepSort and SSD-Sort,MOTA,which are popular in recent years.

关 键 词：遮挡感知 特征融合 运动特征 方向直方图特征 行人检测 行人跟踪 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]