The invention discloses an elevator passenger number detection method based on deep learning, and the method comprises the steps: firstly carrying out the framing of passenger targets in an elevator video frame through employing an elevator target classification training based on deep learning and a coordinate calibration tool; secondly, performing deep learning network optimization; performing multi-level feature map fusion on the neural network, combining high-level features and low-level features, stacking the features in different Channels, correcting a target area, adjusting parameters ofa punching constant and a weight attenuation coefficient, and stopping training when loss does not drop any more or drops extremely slowly to obtain a corresponding target detection weight file; andfinally, inputting the position information of the weight file, the name file and the configuration file obtained by deep learning, generating a dynamic link library, inputting the elevator video in avideo frame format, running a detection file, and outputting the real-time passenger number of the elevator. The passenger number of the elevator is accurately detected, meanwhile, the accuracy of the neural network is improved through parameter optimization, and extremely high efficiency and accuracy are achieved.
一种基于深度学习的电梯乘客数检测方法,首先使用基于深度学习的电梯目标分类训练,采用坐标标定工具,将电梯视频帧中的乘客目标进行框定;其次,进行深度学习网络优化,对神经网络做多级特征图融合,将高层特征与低层特征结合起来堆积在不同的Channel中,矫正目标区域,对冲量常数、权值衰减系数参数作出调整,当loss不再下降或下降极慢时就停止训练,得到相应的目标检测权重文件;最后,输入深度学习所得的权重文件、名称文件与配置文件位置信息,生成动态链接库,将电梯视频以视频帧格式输入,运行检测文件输出电梯实时乘客数。本发明准确检测电梯的乘客数,同时通过参数的优化提高神经网络的准确性,拥有极高的效率与准确性。
Elevator passenger number detection method based on deep learning
一种基于深度学习的电梯乘客数检测方法
2020-06-30
Patent
Electronic Resource
Chinese
Elevator car passenger behavior recognition method based on deep learning
European Patent Office | 2022
|