Elevator Passenger Abnormal Behavior Recognition Method Based

on Digital Twin

Xiaolin Li and Yaohui Song

School of Communications and Information Engineering, Chongqing University of Posts and Telecommunications,

Chongqing 400065, China

Keywords: Digital Twins, Posture Recognition, Bone Extraction, YOLOv7-OpenPose.

Abstract: Aiming at the lack of abnormal behavior dataset and scarcity of samples of elevator passengers, a method

based on digital twin is proposed by this paper to build a vertical elevator passenger abnormal behavior

detection platform and realize the virtual and real mapping of elevator operation status and passenger

behavior. The digital twin scene is combined with the theory of human behavior modeling to enhance the

abnormal behavior of passengers and provide sufficient abnormal behaviour data sources. In order to solve

the problem of confusion between passengers and car background caused by the small range of elevator

monitoring and reduce the accuracy of feature extraction, YOLOv7-OpenPose is used by this paper to

obtain human bone features, which improves the recognition accuracy on the premise of ensuring the

recognition speed, and realizes the rapid recognition of passengers' abnormal behaviors fused with twin

data. Experimental results show that the proposed method not only demonstrates the feasibility, efficiency

and security of digital twin technology in the creation of abnormal data, but also reflects the superiority of

the improved algorithm in pose recognition.

1 INTRODUCTION

Elevator safety is closely related to public safety,

according to the public data of the State

Administration for Market Regulation on national

special equipment, by the end of 2022, the number

of elevators in China has reached 9.6446 million.

Due to the small and closed space of the elevator

car, the passenger dynamics cannot be controlled,

and a series of safety problems such as door

opening, falling, and blocking of cameras are prone

to occur, so the research on the abnormal behavior

of elevator passengers is of great significance.

At present, there is an endless stream of research

on the abnormal behavior of elevator passengers. Lv

et al. (Lv, 2021) used the YOLOv3 algorithm and

the AlphaPose algorithm to collect and extract

elevator passenger behaviors, and used SVM and

neural networks to classify different abnormal

behaviors. Shi et al. (Shi, 2021) used the OpenPose

algorithm to extract the key nodes of the key frame

sequence, obtain the spatial information of the target

behavior, and identify the human behavior in the

elevator car. Feng et al. (Feng, 2021) proposed a

detection method of machine vision and multi-

feature fusion for elevator passenger falls. Wang et

al. (Wang, 2018) used Lucas-Kanade optical flow to

design a passenger abnormal behavior detection

system in the elevator car based on video

recognition technology. Reinsalu et al. (Robal, 2023)

proposed a method for identifying unsafe behaviors

of car passengers based on deep learning and a fault

warning method for brakes and control cabinets

based on infrared spectrum analysis to identify

abnormal behaviors online. Yu and Sun et al. (Yu,

2020), (Sun, 2019) proposed abnormal behavior

detection models based on optical flow method and

angular kinetic energy for the fighting behavior in

the elevator car, respectively.

However, in the process of model training,

testing, and validation, the datasets in the above

studies are insufficient. The traditional elevator

abnormal behavior sample data came from real

human movements and the construction of human

physical models (Yuan, 2020), (Fuller, 2023),

respectively. For datasets created with real human

actions, the execution process has security risks and

requires high time and equipment costs. For the

construction of human body physical models, the

model has fixed actions and lacks real-time and

Song, Y. and Li, X.

Elevator Passenger Abnormal Behavior Recognition Method Based on Digital Twin.

DOI: 10.5220/0012887900004536

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Data Mining, E-Learning, and Information Systems (DMEIS 2024), pages 87-93

ISBN: 978-989-758-715-3

flexibility. The emergence of digital twins provides

a solution to the above problems. The interaction

between the digital twin model and the digital twin

data (Saravanan, 2022) can not only realize the

driving and updating of the digital twin model, but

also support the storage, update and fusion of the

digital twin data. On the one hand, it can realize the

real human-computer interaction process (Amara,

2023), so as to realize the real-time monitoring of

the scene in the elevator. And on the other hand, it

can use the high-fidelity twin scene to reproduce the

abnormal behavior of various passengers, which can

not only ensure the safety of personnel and property

and save equipment costs, but also provide sufficient

data sets for model testing and verification for the

expansion of sample data.

In summary, a method based on digital twins is

proposed by this paper to establish a vertical

elevator passenger behavior detection model. This

method builds a digital twin of elevator passenger

behavior monitoring and a human behavior twin

model, and enhance the abnormal behavior data of

elevator passengers. It provides a large amount of

twin data for the identification of abnormal

behaviors of elevator passengers. For the recognition

of abnormal behaviors, this paper uses the YOLOv7-

OpenPose algorithm to extract features and

iteratively learn from the twin data, so as to realize

the efficient recognition of passengers' abnormal

behaviors.

2 CONSTRUTION OF VERTICAL

ELEVATOR PASSENGER

BEHAVIOR MONITTORING

PLATFORM

2.1 Design of a Digital Twin Framework

for Elevator Passenger Behavior

Monitoring

Referring to the five-dimensional model architecture

of digital twin proposed by He et al. (He, 2020), this

paper builds a framework for the digital twin vertical

elevator passenger behavior monitoring platform, as

shown in Figure 1, which is divided into four layers,

namely the physical layer, the interaction layer, the

virtual layer and the application layer.

(1) Physical layer: The physical layer contains

all kinds of hardware physical entities for elevator

operation, and the main hardware equipment

includes elevator car, camera in the car, elevator

attitude sensor, etc., which are combined into scene

monitoring entities, and ordinary sensors are

responsible for collecting elevators static data, such

as geometric dimensions, car status, etc., attitude

sensors are responsible for collecting passenger

behavior information.

Monitoring and

management

Elevator

Elevator Attitude

Sensor

Camera

Video Analysis

Server (Attitude

Monitoring Port)

Elevator Control

Cabinet (Elevator

Physical Data)

Physical

layer

Interaction

layer

Virtual

layer

Twin data

Model of Abnormal

Passenger Behavior

Elevator status

model

Application

layer

Visual

interface

Statistics

Alert

notifications

Perceptual

access

Algorithm

deployment

Data-driven

Data

support

Figure 1: Elevator passenger behavior monitoring digital

twin framework.

(2) Interaction layer: The interaction layer is

mainly composed of elevator control cabinet and

video analysis server. As the connection layer

between the physical entity and the twin, the

interaction layer can not only use the edge controller

to convey the instructions of the elevator system, but

also upload the simulation data to the video analysis

server for model training, and at the same time

transmit all kinds of collected data to the virtual

layer to drive the call of the twin model and the

interaction of the twin data.

(3) Virtual layer: As the data combination of

physical entities, the virtual layer contains all the

space, attributes and management data of the

physical layer, and stores, transmits, expresses and

deepens them, mainly by means of new technologies

such as digital modeling, Internet of Things, and

artificial intelligence. The twin model is mainly

composed of an elevator state model and a human

behavior model. With the help of virtual and real

interaction technology, the elevator state model

integrates real-time information of elevator

equipment on the basis of the digital model to realize

the real-time monitoring of the operation status of

elevator equipment. The human behavior model uses

video data to realize the virtual and real mapping of

real passenger human behavior, and at the same

time, the digital twin character model creates a large

amount of abnormal behavior data. The twin data

comes from the physical layer and the virtual layer,

the data of the physical layer is used to drive the

twin model, and the data supply of the virtual layer

is called by the layer and the interaction layer.

DMEIS 2024 - The International Conference on Data Mining, E-Learning, and Information Systems

(4) Application layer: The application layer

directly serves the staff, including the visualization

of the twin interface and the operation functions of

safety warning, and displays the elevator status and

the elevator passenger status in real time, so as to

realize the three-dimensional dynamic supervision of

elevator equipment and passenger behavior. It

mainly uses UI design and three-dimensional

visualization technology, and connects to the

communication system at the same time, so that

elevator passengers can take rescue actions in the

first time in case of safety accidents.

2.2 Construction of Digital Twin Scene

for Passenger Behavior Monitoring

The vertical elevator passenger behavior monitoring

digital twin platform was built in Unreal Engine 5

(UE5 Unreal Engine), as shown in Figure 2, which is

divided into three steps: geometric model drawing,

twin scene construction, and virtual and real

mapping.

Visual service platform

Device state

model

Human

behavior

model

Digital twi n g eome tr y

UE5 Unreal Engine(Actors,

Charts, Fonts, etc)

Blueprint, C++, C# compilation

Physical Entities: Elevators,

Passengers

Twin data

mapping

Figure 2: Construction of digital twin scene for elevator

passenger behavior monitoring.

2.2.1 Geometric Model Drawing

The geometric model is drawn according to the

physical entity behavior rules and each system, and

the elevator model is drawn by SolidWorks,

including eight systems: drag system, car, weight

balance system, power system, door system and

safety protection. The character model was drawn in

Maya and had to be drawn with a skeletal mesh to

reflect the real human behavior.

2.2.2 Twin Scene Building

Import the above geometry into UE5 and assemble it

according to the inclusion relationship in the

physical world. When building a scene, pay attention

to redefining material properties, hierarchical

relationships, motion constraints, and so on. Ensure

that the movement of the twin in the twin scene is

consistent with the physical world.

2.2.3 Virtual-Real Data Mapping

Twin data is the key to real-time mapping of

physical entities and twins. In this paper, the

elevator status data is obtained through the RS485

communication bus between the external interface

board and the elevator control cabinet, and stored in

the corresponding MySQL data for driving the twin.

Passenger behavior data includes video data and

passenger attitude information, wherein passenger

attitude information is obtained by using the

YOLOv7-OpenPose algorithm proposed in this

paper on the video analysis server, and sent to the

client through TCP network protocol and Socket

communication, and the client can realize the

construction of visualization platform services by

using C++ and BluePrint to parse and distribute the

input data.

3 IDENTIFICATION OF

ABNORMAL PASSENGER

BEHAVIOR FUSED WITH

TWIN DATA

3.1 Passenger Abnormal Behavior

Modelling and Data Augmentation

Abstracting the human skeleton into a collection of

several bones and joint points can not only reduce

the complexity of constructing a complete human

behavior model, but also the human bone structure is

similar, and the human behavior can also be

described using the position information of these

joints. As shown in Figure 3, The traditional 3D

coordinate description method makes the position

parameters of each joint point independent of each

other during the movement. It violates the constraint

that the skeleton length of the mannequin remains

unchanged, so this paper directly uses a hierarchical

method to equivalence the mannequin to the joint

tree shown in the figure above. The root node of the

tree is used as the geometric center of the human

skeleton model to control the overall displacement

and direction of the model. The remaining sub-nodes

are indirectly or directly connected to the root node

to assist in the determination of the posture of the

mannequin and the presentation of the movement

process (Han, 2020).

Elevator Passenger Abnormal Behavior Recognition Method Based on Digital Twin

Root

R_Hip

Pelvis

L_Hip

R_Knce

R_Ankle

L_Knce

L_Ankle

Chest

Neck

R_shouder

L_shouder

Head

R_Elbow L_Elbow

R_Wrist R_Wrist

Figure 3: Human joint tree.

As shown in Figure 4, after equating the human

skeleton to a joint tree, the movement of the whole

human body can be described in detail by means of

forward kinematics and inverse kinematics

(Moharkan, 2023).

Move bone

points to

generate

keyframe

actions

Bone retargeting

Solve for

refined bone

data between

adjacent

keyframes

Inverse kinematics

Generate

multiple sets

of action

sequences

Positive kinematics

The state

machine

integrates

multiple sets

of actions

into a

character

State integration

Twin scenes

reproduce the

action

Event triggering

Figure 4: Digital twin-based passenger anomaly modelling

process.

Among them, after the displacement and rotation

of each joint relative to its parent joint are known,

the posture of the whole human body after

movement can be solved by using positive

kinematics. With the initial and ending human

posture known, inverse kinematics can be used to

solve for specific changes in each joint during the

movement process.

Through forward and reverse kinematic behavior

modeling, various dangerous posture actions can be

created for the twin model to operate, because the

abnormal behavior recognition research of machine

learning requires a large amount of video image

data. In the process of acquisition, the frame rate

output of the CameraActor is set to 30frame/s, the

resolution is set to 1920×1080, the image output

format is video sequence (avi), the angle is set to the

horizontal plane of the parallel car roof, and the four

virtual cameras are recorded at the same time

through the control command. Data enhancement is

carried out for abnormal behaviors such as jumping,

falling, kicking, and hand-picking doors in the

elevator.

3.2 Human Bone Detection Based on

YOLOv7-OpenPose

Aiming at the lack of abnormal behavior cases and

insufficient datasets in real life, the abnormal

behavior modeling of passengers based on digital

twins can provide a large number of training and test

samples for abnormal posture and behavior. In this

paper, we propose a passenger abnormal behavior

recognition model fused with twin data, and the

overall framework diagram of the model is shown in

Figure 5.

Input

Real video

Twin video

Image

sequence

Skeleton extraction by Yolo-OpenPose

VGG-19

Yolo

Backbone

Branch1，Stage_1

Branch1，Stage_t

Branch2，Stage_1

Branch2，Stage_t

⊕⊕

Feature

Decoupled into2D skeleton，

Preprocessing feature and get

Skeleton sequence

-0.2

0.2

0.4

-0.4

-0.6

Figure 5: Bone extraction of passenger abnormal behabior

fused with twin data.

The video data of passenger abnormal behavior

collected in practice and the twin data of passenger

abnormal behavior constructed in 2.1 were

processed into image sequences, and the skeletal

posture of the human body was extracted by the

YOLOv7-OpenPose algorithm.

OpenPose, as a classical pose extraction

algorithm, has a multi-stage, supervised

convolutional neural network (Moharkan, 2023),

which is widely used in human behavior recognition,

and its overall structure is shown in B of Figure 5.

First, the feature map is extracted through the

backbone

. Secondly, it inputs this map into

Branch1 and Branch2 of the first stage Stage_1. And

then it obtains the 2D confidence map of the joint

point

S and the partial affinity domain

respectively. Finally, it makes

−

and

−

the input of the next stage Stage_

t ( 2t ≥ ), and

perform multiple iterations. The inputs and outputs

of Branch1 and Branch2 at different stages are

shown in Eq. (1).

(, , ), 2

tt tt

SFSLt

LFSLt

−−



=∀≥





=∀≥





(1)

, where

and

, represent the network

relationship between the 2D confidence graph and

DMEIS 2024 - The International Conference on Data Mining, E-Learning, and Information Systems

partial affinity domain of the t stage and the feature

map, 2D confidence graph and partial affinity

domain of the stage

1t − , respectively. Through the

continuous iteration of the above multi-stage

convolutional neural network, the more joint points

of the human body are obtained. For the position

d and

d of any two joint points, the confidence

degree of the joint point pair needs to be

characterized by calculating the linear integral of the

partial affinity domain by Eq. (2). At last, it selects

the joint with the highest confidence to complete the

splicing, and gets the skeletal posture of the whole

human body.

(())

() (1 )

ELpu d

pu ud ud

−



=×



−





=− +





(2)

, where

L represents a partial affinity domain for a

pair of joint points.

Although the OpenPose algorithm has the

advantages of fast recognition speed and high

accuracy in large scenarios. However, the scope of

elevator monitoring is small, which is easy to cause

confusion between passengers and the car

background and reduce the accuracy of feature

extraction. So it is necessary to consider more

dependencies between deep and shallow networks to

capture more details and textures. VGG-19 (Wen,

2019) is the feature extraction network of the

OpenPose algorithm, and the continuous iteration of

the convolutional network makes the significance of

the feature map continuously decrease. Therefore, in

order to better extract the skeletal features of

passengers in the car environment, this paper

improves the traditional VGG-19 and proposes a

skeletal point detection algorithm for elevator

passengers based on YOLOv7-OpenPose. The

backbone network structure of the algorithm is

shown in Figure 6.

This structure uses the YOLOv7 object detection

model to extract abnormal passenger targets. The

YOLOv7 series uses both object detection and

object classification, and its object detection layer of

the 79th layer 13*13 detection network and the

object detection layer of the 91-layer 26*26

detection network are both convolutional layers, and

only feature extraction is performed on the target.

The specific position, bounding box and confidence

level of the target are obtained through the object

detection network of YOLOv7, which is placed

before the third Maxpool layer in the VGG-19

network as input. And then the output tensor of the

whole object detection is spliced with the output

tensor of the original Maxpool layer as the input

tensor of the next layer and the bone extraction is

continued.

4 EXPERIMENTAL RESULTS

AND ANALYSIS

4.1 Experimental Environment

This experiment adopts Unreal Engine 5.3.2 under

Windows 10 operating system, TensorFolw

framework, Core i9-10980XE processor, 128G

Yolov7 Object detection model

Conv3-64

Conv3-128

Conv3-256

Conv3-512

Maxpool

CSPNet

Upsampling Layer

OpenPose model OpenPose model

Figure 6: YOLOv7-openpose algorithm backbone.

Elevator Passenger Abnormal Behavior Recognition Method Based on Digital Twin

memory, Core i7-11400F processor, 32G memory,

Nvidia RTX 3080 graphics processor and Nvidia

RTX3060 graphics processor.

4.2 Dataset

The experiments were performed on a custom

elevator passenger abnormal behavior dataset and a

COCO2017 dataset. Among them, the COCO2017

dataset is used to test the YOLOv7-OpenPose bone

point detection algorithm in this paper. The dataset

contains a total of 163957 images of the training,

validation, and test sets, and each image is labeled

with 18 bone points of the human body.

By capturing and merging the videos in the real

scene and the twin scene at a rate of 30 frames per

second, the custom passenger abnormal behavior

dataset shown in Table 1 is formed.

Table 1: Customize the Passenger Abnormal Behavior

Dataset.

Action type

The amount of

real data

The amount of

virtual data

Total

Caper 28 2023 20510

Hand chop 47 2028 2075

Kicks 66 2047 2113

Fall 20 2011 2031

Total 161 8109 8270

4.3 Human Skeletal Point Detection

Analysis

Human skeleton detection based on YOLOv7-

OpenPose is not only the key link to obtain

passenger posture information in the real

environment, but also the core content of obtaining

passenger abnormal behavior dataset. For the

abnormal behaviors of various passengers

constructed in the twin scene, the skeletal point

detection results are shown in Figure 7.

4.4 Comparison of the Performance of

the YOLOv7-OpenPose Algorithm

In order to verify the accuracy and real-time

performance of the YOLOv7-OpenPose bone

detection algorithm, the performance of the text was

compared with the classical OpenPose algorithm on

the COCO dataset, and the performance of different

series of YOLO algorithms was analyzed. The

experimental results are shown in Table 2.

Caper Hand chop Kicks Fall

Figure 7: Bone diagnosis of abnormal passenger behavior

based on twin data.

Table 2: Comparison of the performance of different

networks on the COCO test set.

The type of

algorithm

OpenPose

YOLOv5

OpenPose

YOLOv7

OpenPose

YOLOv8

AP 0.701 0.728 0.734 0.741

0.854 0.871 0.882 0.845

0.711 0.722 0.736 0.709

0.695 0.714 0.719 0.725

0.774 0.776 0.779 0.785

Time-

consumed

21341 23448 22154 25243

Where AP is the mean index of average accuracy.

and AP

represent the joint point predictors

with thresholds of 0.50 and 0.75, respectively, AP

and AP

represent the predictors of small and large

human body size, respectively. Time-consumed is

the test time consumed by the algorithm on the

experimental platform in this paper, which is used to

characterize the algorithm complexity. Comparing

the accuracy and time complexity of each model in

Table II, it can be seen that the AP

and AP

prediction indicators of YOLOv7-OpenPose perform

the best among the above algorithms. Compared

with the original model OpenPose, the accuracy

indicators of YOLOv7-OpenPose are improved and

the amplitude is about 4%, and the introduction of

the YOLOv7 module does not have a great impact

on the time complexity of the overall YOLOv7-

OpenPose algorithm. The average processing time of

the algorithm is still close to that of the original

OpenPose algorithm, which can meet the real-time

requirements.

DMEIS 2024 - The International Conference on Data Mining, E-Learning, and Information Systems

5 CONCLUSIONS

In order to solve the problem of lack of a large numb

er of samples in the research on abnormal behavior r

ecognition based on machine learning caused by the

scarcity of abnormal behavior data of elevator passe

ngers, this paper builds a digital twin scenario for m

onitoring abnormal behavior of passengers in vertica

l elevators. Based on the human skeleton model and

kinematics principle, the abnormal behavior data wa

s constructed, and a total of 8270 twin actions were

provided for abnormal behavior recognition, and fin

ally the improved YOLOv7-OpenPose human skelet

on detection algorithm was used. The experimental r

esults show that the accuracy of the model is improv

ed by about 4% on the basis of the original model O

penPose, and the model does not significantly increa

se the time complexity in terms of real-time, which s

olves the problem of low feature extraction rate caus

ed by the confusion between passengers and car bac

kground and the trade-off between real-time and acc

uracy. The modeling and recognition of abnormal be

haviors proposed in this paper have the characteristic

s of high accuracy, strong real-time performance and

good interactivity. In the future, further research wil

l be carried out on the abnormal behavior of multiple

people in complex scenes in the elevator.

REFERENCES

Lv, X., Wang, S. and Chen, T., 2021. Human gait analysis

method based on sample entropy fusion AlphaPose alg

orithm. 2021 33rd Chinese Control and Decision Conf

erence (CCDC). pp.1543-1547. IEEE. doi: 10.1109/C

CDC52312.2021.9602427.

Shi, D., Zhu, W. and Cheng, R., 2022. Human Fall Detecti

on Algorithm Based on YoloX-s and Lightweight Ope

nPose. 2022 15th International Conf. on Advanced Co

mputer Theory and Engineering (ICACTE). pp. 23-28.

IEEE. doi: 10.1109/CCDC52312.2021.9602427.

Feng, S., Niu, K. and Liang, Y., 2021. Research on elevato

r intelligent monitoring and grading warning system. I

EEE International Conf. on Computer Science, Electr

onic Information Engineering and Intelligent Control

Technology (CEI). pp. 145-148. IEEE. doi: 10.1109/C

EI52496.2021.9574579.

Wang, Z., Yang, X., 2018. Moving target detection and tra

cking based on Pyramid Lucas-Kanade optical flow.20

18 IEEE 3rd International Conference on Image, Visio

n and Computing (ICIVC). pp. 66-69. IEEE. doi: 10.11

09/ICIVC.2018.8492786.

Reinsalu, U., Robal, T., 2023. A Touch-Free Service Butto

n for Smart Elevator Operation with Dynamic QR-cod

e Generation. 2023 12th Mediterranean Conference on

Embedded Computing (MECO). pp. 1-4. IEEE. doi: 1

0.1109/MECO58584.2023.10155102.

Yu, L., Wang, Y. and Li, J., 2020. Algorithm of Detecting

Abnormal Behavior Based on Optical Flow Method. 2

020 2nd International Conference on Information Tec

hnology and Computer Application (ITCA). pp. 191-1

93. IEEE. doi: 10.1109/ITCA52113.2020.00047.

Sun, Z., Xu, B. and Lu, M., 2019. A real-time video survei

llance and state detection approach for elevator cabs. 2

019 International Conference on Control, Automation

and Information Sciences (ICCAIS). pp. 1-6. IEEE. doi

: 10.1109/ICCAIS46528.2019.9074707.

Yuan, R., Zhang, Z. and Song, P., 2020. Construction of vi

rtual video scene and its visualization during sports tra

ining. IEEE Access. vol. 8, pp. 124999-125012. IEEE.

doi: 10.1109/ACCESS.2020.3007897.

Kyrollos, D., Fuller, A. and Greenwood, K., 2023. Under t

he cover infant pose estimation using multimodal data.

IEEE Transactions on Instrumentation and Measurem

ent. vol. 72, no. 5007212, pp. 1-12. IEEE. doi: 10.110

9/TIM.2023.3244220.

Saravanan, S., Muthusenthil, B. and Gurusubramani, S., 2

022. A review of digital twin leveraging technology, c

oncepts, tools and industrial applications. 2022 1st Inte

rnational Conf. on Computational Science and Techno

logy (ICCST). pp. 742-748. IEEE. doi: 10.1109/ICCST

55948.2022.10040359.

Amara, K., Kerdjidj, O. and Ramzan, N., 2023. Emotion R

ecognition for Affective human digital twin by means

of virtual reality enabling technologies. IEEE Access. vol.

11,pp.74216-74227,doi:10.1109/ACCESS.2023.3285398

He, Z., Yuan, T. and Li, X., 2022. Research on Data Transf

er Scheme of Virtual-Real Interactive in Five-Dimensio

nal Digital Twin System. 2022 IEEE 8th International

Conf. on Computer and Communications (ICCC). pp. 23

39-2344. IEEE. doi:10.1109/ICCC56324.2022.10066033.

Han, B., Jiao, Y. and Liu, G., 2020. Kinematic & dynamic

models of human lower extremity during the gait cycl

e. 2020 6th International Conference on Control, Auto

mation and Robotics (ICCAR). pp. 568-573. IEEE. do

i: 10.1109/ICCAR49639.2020.9108090.

Moharkan, M. J., 2023. Comprehensive Survey on Human

Motion Tracking System. 2023 2nd International Con

ference on Automation, Computing and Renewable Sys

tems (ICACRS). pp. 1380-1387. IEEE. doi: 10.1109/I

CACRS58579.2023.10405344.

Wen, L., Li, X., 2019. A new transfer learning based on V

GG-19 network for fault diagnosis. 2019 IEEE 23rd In

ternational Conf. on Computer Supported cooperative

work in design (CSCWD). pp. 205-209. IEEE. doi: 10.

1109/CSCWD.2019.8791884.

Elevator Passenger Abnormal Behavior Recognition Method Based on Digital Twin