嵌入中心点预测模块的Yolov3遮挡人员检测网络  

Embedded Center Prediction Module of Yolov3 Occlusion Human Detection Network

作　　者：梁煜 李佳豪 张为 孙琦龙[2] Liang Yu;Li Jiahao;Zhang Wei;Sun Qilong(School of Microelectronics,Tianjin University,Tianjin 300072,China;School of Computer Science,Qinghai Nationalities University,Xining 810007,China)

机构地区：[1]天津大学微电子学院,天津300072 [2]青海民族大学计算机学院,西宁810007

出　　处：《天津大学学报（自然科学与工程技术版）》2021年第5期517-525,共9页Journal of Tianjin University：Science and Technology

基　　金：公安部技术研究计划资助项目(2017JSYJC35);青海民族大学理工自然科学重大项目(2019xjz003);新一代人工智能科技重大专项(19ZXNGX0030).

摘　　要：为解决目前实际监控场景下人员检测任务中存在的遮挡问题,提出了一种改进的Yolov3检测网络.首先,针对现有人员检测算法的被检测目标姿态单一且大多是室外直立行人的问题,自建了一个包含16832张样本的多场景人员检测数据集用于对检测网络进行训练和测试,其中包含训练集样本12090张,测试集样本4742张.随后,为了提升网络在遮挡情况下的检测效果,设计了中心点预测模块(CPM)并将其嵌入到原Yolov3网络中3个尺度的输出特征图上,通过该模块首先确定目标的中心位置作为预提取的中心点,随后在此预提取的中心点上对目标的位置和尺寸进行精确的回归.最后,候选框的精确回归中采用广义的交并比指标来构造损失函数进行优化,通过准确地构造候选框和真实目标框的位置关系来提高其回归精度,同时降低损失函数在不同尺度目标下的波动.实验结果表明:优化网络结构和损失函数后的检测网络在测试集上的检测精度提高了2.92%,漏检率下降2.94%,针对实际监控场景下的遮挡情形取得了很好的检测效果,而且对多姿态人员目标的检测结果具有很好的鲁棒性;同时检测速度达到了28帧/s,保证了检测的实时性.另外,在Caltech行人数据库上该网络的漏检率为6.02%,相对于传统的检测网络同样达到了最优的效果,进一步印证了网络在行人检测任务上的优越性.To solve the occlusion problem in the current human detection task in actual monitoring scenarios,an improved Yolov3 detection network was proposed.First,in view of the problem that the detected target posture of the existing human detection algorithms is that of single,mostly outdoor,upright pedestrians,a multi-scene human detection dataset(MHDD)containing 16832 samples was self-built for training and testing the network,which included 12090 samples in the training set and 4742 samples in the test set.Then,to improve the detection effect of the network in the case of occlusion,the center prediction module(CPM)was designed and embedded into the threescale output feature map of the original Yolov3 network.This module first determined the center position of the target as the pre-extracted center point,and then the location and size of the target were accurately regressed on it.Finally,in the accurate regression of the candidate boxes,the GIoU(generalized intersection over union)was used to construct the loss function for optimization,and the regression accuracy was improved by accurately constructing the position relationship between the candidate boxes and real target boxes,which also reduced the fluctuation of the loss function under different scale targets. The experimental results show that the detection accuracy of the detection networkon the test set after optimizing the network structure and the loss function is increased by 2.92%,and the misseddetection rate is decreased by 2.94%. The network achieves a good detection effect for the occlusion situation in actualmonitoring scenarios,and it has good robustness for the detection results of multi-pose human targets. At the sametime,the detection speed reaches 28 frames per second,ensuring real-time detection. In addition,the missed detectionrate of the network on the Caltech pedestrian database is 6.02%,which also achieves better results than those ofthe traditional detection networks,further confirming the superiority of the network in pedestrian detection tas

关 键 词：计算机视觉 视频监控 卷积神经网络 人员检测 人员遮挡 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]