摘要
由于光照变化、乘客拥挤和站外噪声干扰大等问题,现今地铁进站客流人脸检测技术精度较低。为提高人脸检测精度,本文在YOLO2轻量级目标检测算法Tiny YOLO2原有网络结构基础上,首先利用不同数目的1×1卷积层对特征图进行压缩,然后将特征图尺寸重新调整到统一大小进行级联,得到高维特征图。缩减网络最后几层卷积核数量,用1×1卷积层替换原始网络的3×3卷积层,得到更深而且更窄的人脸检测网络。改进后的网络先后在Wider Face数据集和地铁进站客流数据集上进行训练,得到最终的人脸检测模型。加载训练好的人脸检测模型对随机选取的300幅站外乘客图片进行测试。测试结果表明:本文算法相比Tiny YOLO2原始人脸检测算法,召回率提高4.2%,单幅图片检测速度提高6.5%。同时在广泛使用的人脸检测算法评测数据集FDDB上进行测试,在误检数目为200的情况下,人脸检测准确率相比Tiny YOLO2平均提高5%,比SSD检测算法提高2%,而且本文算法能够在检测速度和精度之间取得较好的平衡,有较好的泛化性。
Due to the problems of illumination change, passenger congestion and large noise interference outside the station, the accuracy of face detection technology for subway passenger flow into station is low nowadays. In order to improve the accuracy of face detection, based on the original network structure of YOLO2 lightweight target detection algorithm Tiny YOLO2, this paper firstly compresses feature maps with different number of 1×1 convolution layers, and then adjusts the size of feature maps to a unified size for cascading to obtain high-dimensional feature maps. We reduce the number of convolution kernels in the last few layers of the network, replace the 3×3 convolution layer of original network with 1×1 convolutional layer to get a deeper and narrower face detection network. The improved network has been trained on the Wider Face dataset and the subway inbound passenger flow dataset to obtain the final face detection model. The trained face detection model is loaded to test 300 randomly selected images of passengers outside the station. The test results show that compared with the Tiny YOLO2 original face detection algorithm, the recall rate is increased by 4.2%, and the detection speed of single image is increased by 6.5%. At the same time, it is tested on the FDDB dataset which is widely used for face detection algorithm evaluation. When the number of false detections is 200, the accuracy of face detection is 5% higher than that of Tiny YOLO 2 and 2% higher than that of SSD. Moreover, this algorithm can achieve a good balance between detection speed and accuracy, and has better generalization.
作者
周慧娟
张强
刘羽
王旭阳
柳颖
ZHOU Hui-juan;ZHANG Qiang;LIU Yu;WANG Xu-yang;LIU Ying(Beijing Key Lab of Urban Intelligent Traffic Control Technology,North China University of Technology,Beijing 100144,China)
出处
《计算机与现代化》
2019年第10期76-82,共7页
Computer and Modernization
基金
国家重点研发计划资助项目(2016YFB1200402)
