Recognition Of Passenmers’ Abaormal Behavior In Elevator Car Rased On Convolutional Network

With the rapid development of urbanization,the application of elevators in urban high-rise buildings is becoming increasingly widespread,the detection of the safety of passenger behavior in elevator cars has also become the focus of people’s attention.However,traditional elevator monitoring video technology can only realize simple recording and playback functions..This forces us to manually check the abnormal behaviors in the elevator monitoring video,which is not conducive to the automatic recognition of abnormal behavior in the elevator car.Based on this background,this paper proposes a reconstructed three-dimensional convolution optical flow feature fusion method to perform feature learning on the surveillance video in the elevator car to realize the function of automatically detecting abnormal behavior of elevator car passengers.This paper,taking the behavior of passengers in the elevator car as the research object,First,it analyzes several common abnormal behavior states and normal behavior states in the elevator car to form an elevator car passenger behavior data sample set,and analyze the image characteristics of the collected video data to determine whether there is a moving target in the elevator.Combined with the actual environment,three common moving target detection algorithms are evaluated and selected from the detection rate,missed detection rate and detection speed.Secondly,select a 3D convolutional network which can self-extract the features of video frames and images,and use a method for identifying abnormal behavior based on reconstructed 3D convolutions.Fitting 3D convolution with 2D convolution and ID convolution makes the network easier to optimize.At the same time,in order to strengthen the influence of behavior information on the recognition result,optical flow characteristics are added to fuse the reconstructed convolution network.Then,the effects of different input frame lengths and different optimization methods on the recognition accuracy of the three models are studied.The experimental results can be analyzed.Reconstructed convolutional network has certain advantages in recognition accuracy compared to 3D convolutional network due to the increase of the number of layers and nonlinearity of the network,and the recognition results of reconstructed convolutional networks with optical flow characteristics are even more optimal.The respectively optimal recognition accuracy of the three models can reach 75.5%,85.26%and 89.69%.The final reconstructed convolution optical flow feature fusion network is optimized by two methods of learning rate adjustment and transfer learning.The optimized recognition accuracy rate can be reach 94.20%.At last,Pytorch was used as the deep learning research framework,combined with the front-end and back-end interactive framework to complete the development of abnormal behavior detection system And,related functions of the system are introduced and system-related tests are finally performed.Test results show that the method proposed in this paper can effectively detect abnormal behaviors in elevator cars,and it has certain application value.