梁 煜 ,李佳豪 ,张 为 ,孙琦龙
AuthorsHTML:梁 煜 ¹ ，李佳豪 ¹ ，张 为 ¹ ，孙琦龙 ²
AuthorsListE:Liang Yu ,Li Jiahao,Zhang Wei,Sun Qilong
AuthorsHTMLE:Liang Yu ¹，Li Jiahao¹，Zhang Wei¹，Sun Qilong²
Unit: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

Unit_EngLish:1. School of Microelectronics，Tianjin University，Tianjin 300072，China；
2. School of Computer Science，Qinghai Nationalities University，Xining 810007，China

Abstract_Chinese: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 16 832 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 12 090 张，测试集样本 4 742 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．
Abstract_English: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 16 832 samples was self-built for training and testing the network，which included 12 090 samples in the training set and 4 742 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 three scale 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 con\u0002struct 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 lossfunction under different scale targets. The experimental results show that the detection accuracy of the detection network on the test set after optimizing the network structure and the loss function is increased by 2.92%，and the missed detection rate is decreased by 2.94%. The network achieves a good detection effect for the occlusion situation in actual monitoring scenarios，and it has good robustness for the detection results of multi-pose human targets. At the same time，the detection speed reaches 28 frames per second，ensuring real-time detection. In addition，the missed detection rate of the network on the Caltech pedestrian database is 6.02%，which also achieves better results than those of the traditional detection networks，further confirming the superiority of the network in pedestrian detection tasks.
Keyword_Chinese:计算机视觉；视频监控；卷积神经网络；人员检测；人员遮挡
Keywords_English:computer vision；video surveillance；convolutional neural network；human detection；human occlusion

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6631