摘要
在轨道交通领域,需要实时监控不同区域的行人位置分布,以解决乘客摔倒,排队时占用下客区等异常情况。现如今设计一种更加轻量化同时不会大幅度降低检测效果的网络成为深度学习行人实时检测中十分具有现实意义的问题。论文针对现有的目标检测算法对计算资源消耗较大的问题,设计了一种基于卷积、可分离卷积与Inception结构的轻量卷积神经网络——LPDNet。该网络在现有的YOLOV3目标检测网络的基础上,只保留了32倍下采样分支,同时加入了Inception-V3-3网络模块,并改进了原有的损失函数。论文设计的网络结构在场景多样的VOC数据集上进行行人检测实验,结果与原网络模型检测的准确率相差不到3%,在地铁相对单体固定的场景下,检测效果的差距仅为0.54%。实验结果表明,新设计出的算法能够在检测效果下降很小的情况下,大大减小计算所需资源,加快推理速度,减小模型大小。
In the field of rail transit,it is necessary to monitor location distribution of pedestrians in different areas in real time to solve the abnormal situations such as passengers falling down and occupying the passenger area when queuing.At present,the design of a more lightweight network without greatly reducing the detection effect has becomes a very practical problem in the real-time detection of pedestrians in deep learning field.Aiming at the problem that the existing object detection algorithm consumes a lot of computing resources,a lightweight convolution neural network LPDNet based on convolution,separable convolution and Inception structure is designed in this paper.Based on the existing YOLOV3 object detection network,this network only retains 32 times the subsampling branch,adds Inception-V3-3 network module,and improves the original loss function.The designed network structure has less than 3% difference in detection accuracy on the VOC data set with various scenes,and the difference in detection effect is only 0.54% in the fixed scene of the subway.The experimental results show that the new algorithm can greatly reduce the computational resources,accelerate inference speed and reduce the size of the model when the detection effect drops very little.
作者
司广字
刘光杰
陆斌
焦科杰
王政军
SI Guangzi;LIU Guangjie;LU Bin;JIAO Kejie;WANG Zhengjun(School of Electronics and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044;Nanjing Panda Information Industry Co.,Ltd.,Nanjing 210038;Hikvision,Nanjing 211106)
出处
《计算机与数字工程》
2023年第2期336-341,371,共7页
Computer & Digital Engineering
基金
国家自然科学基金项目(编号:U1836104,61772281,61702235,61801073,61931004,62072250)
江苏省研究生科研与实践创新计划项目(编号:KYCX20_0974)
教育部人文社科基金项目(编号:19YJA630061)资助。
关键词
轨道交通
行人检测
卷积神经网络
轻量化模型
rail transit
pedestrian detection
convolutional neural network
lightweight model
