摘要
针对公交客流数据获取现状不稳定、处理方法陈旧及无法实时调度等问题,本文结合图像处理和深度学习方法,设计了公交客流检测系统,重点对乘客头部的特征提取算法进行研究。该系统将工业摄像头分别置于车厢前端、中端及末端的上侧,以此获得车内乘客三通道图像,在Raspberry Pi树莓派中移植TensorFlow深度学习框架,由此搭建卷积神经网络CNN的模型,以此模型提取车厢内乘客头部特征,再采用综合梯度下降算法优化学习速率和三通道数据融合技术判断车内拥挤情况,从而保证模型的鲁棒性。实验结果表明:将识别结果输入SPP-Net网络结构中,上述算法识别准确率为87.23%,较传统卷积神经网络提高9.11%,收敛速率提高20.77%,其实时的拥挤度判断更具有实用性。
Generally,bus traffic statistics can provide the entire passenger flow data of operating vehicles.Vehicle dispatching management personnel use this data to make statistics,analysis and decision,so as to realize the effective allocation of vehicles and personnel,thus highlighting the advantages of intelligent dispatching.Aiming at the problems of unstable current situation of bus passenger flow data acquisition,outdated processing methods and inability to schedule in real time,a set of bus passenger flow detection system is designed based on image processing and deep learning methods,and the feature extraction algorithm of the passenger head is deeply studied.Industrial cameras in system design are respectively placed on the upper side of the front end,the upper side of the middle end and the upper side of the tail end of the car to obtain three-channel images of passengers in the car.A TensorFlow depth learning framework is transplanted into Raspberry Pi and a convolution neural network CNN model is built to extract the head characteristics of passengers in the car.The learning rate is optimized through a comprehensive gradient descent algorithm,and the congestion in the car is judged by using a three-channel data fusion technology to ensure the robustness of the model.Experiments show that when the recognition results are input into the SPP-Net network structure,the recognition accuracy of the above algorithm is 87.23%,which is 9.11%higher than the traditional convolutional neural network and the convergence rate can be increased by 20.77%,which is more practical than real-time congestion judgment.
作者
张开生
刘泽新
郭碧筱
杨帆
ZHANG Kaisheng;LIU Zexin;GUO Bixiao;YANG Fan(School of Electrical and Information Engineering,Shaanxi University ofScience and Technology,Shaanxi,Xi’an 710021,China)
出处
《石河子大学学报（自然科学版）》
CAS
北大核心
2019年第5期654-660,共7页
Journal of Shihezi University(Natural Science)
基金
国家自然科学基金项目(61601271)
陕西省科技厅社会发展科技攻关项目(2016SF-418)
关键词
客流检测
卷积神经网络CNN
综合梯度下降算法
拥挤情况
passenger flow detection
convolutional neural networkcnn
integrated gradient descent algorithm
congestion
