一种改进RetinaNet的室内人员检测算法  

Indoor human detection algorithm based on the improved retinaNet

作　　者：王璐璐 张为 孙琦龙[3] WANG Lulu;ZHANG Wei;SUN Qilong(School of Electrical Automation and Information Engineering,Tianjin University,Tianjin 300072,China;School of Microelectronics,Tianjin University,Tianjin 300072,China;School of Computer Science,Qinghai Nationalities University,Xining 810007,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072 [2]天津大学微电子学院,天津300072 [3]青海民族大学计算机学院,青海西宁810007

出　　处：《西安电子科技大学学报》2019年第5期69-74,104,共7页Journal of Xidian University

基　　金：国家部委技术研究计划(2017JSYJC35);青海民族大学理工自然科学重大项目(2019xjz003)

摘　　要：由于现有的人员检测算法研究对象主要是室外直立行人,而室内人员姿态多变,且图像拍摄角度与室外行人差别较大,所以使用以往的检测方法得到的效果并不理想。基于此,笔者针对室内人员检测数据集提出了一种高精度检测模型。该模型以RetinaNet网络为基础,在残差网络中引入通道注意力模块,间接实现卷积层的随机失活,增强模型泛化能力;通过维度聚类算法找出锚点的最佳尺寸,并据此找到合适的特征图进行预测。实验表明,这种算法在室内人员检测数据集上检测精度可达99.84%,且在速度和内存占用方面也优于其他算法。Human detection is of great significance in computer vision tasks such as security and human-machine interaction. In this paper, a high-precision detection model based on the indoor human detection dataset(IHDD) is proposed for indoor human detection. As the posture of the indoor staff is changeable and the image shooting angle is quite different from that of outdoor pedestrians, the model we propose makes significant improvement in the field of human detection. In this work, we integrate the Squeeze-and-Excitation module into the residual network to realize the dropout of the convolutional layer to enhance the generalization ability of the model. Meanwhile, dimension clustering is utilized to find the optimal size of anchors and the best feature map to be used in prediction. Experimental results on IHDD demonstrate that the proposed methods can reach a precision of 99.84% and outperform other algorithms in terms of speed and memory usage. It indicates that our method has a certain theoretical and practical value.

关 键 词：机器视觉 卷积神经网络 室内人员检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]