Development of a Realistic Crowd Simulation Environment for

Fine-grained Validation of People Tracking Methods

Preprint, compiled April 27, 2023
ID 1 ID 1
Paweł Foszner , Agnieszka Szczesna
˛ , Luca Ciampi ID 2 , Nicola Messina ID 2 , Adam Cygan3 , Bartosz ∗
Bizoń3 ,
Michał Cogiel ID 4 , Dominik Golba ID 4 , Elżbieta Macioszek ID 5 , and Michał Staniszewski ID 1
1
Department of Computer Graphics, Vision and Digital Systems, Faculty of Automatic Control, Electronics and Computer
Science, Silesian University of Technology, Gliwice, Poland; [email protected]
2
Institute of Information Science and Technologies, National Research Council, Pisa, Italy; [email protected]
3
QSystems.pro sp. z o.o. Mochnackiego 34, 41-907 Bytom, Poland; [email protected]
4
Blees sp. z o.o. Zygmunta Starego 24a/10, 44-100 Gliwice, Poland; [email protected]
5
Department of Transport Systems, Traffic Engineering and Logistics, Faculty of Transport and Aviation Engineering, Silesian
arXiv:2304.13403v1 [cs.CV] 26 Apr 2023

University of Technology, Katowice, Poland; [email protected]

Abstract
Generally, crowd datasets can be collected or generated from real or synthetic sources. Real data is generated
by using infrastructure-based sensors (such as static cameras or other sensors). The use of simulation tools
can significantly reduce the time required to generate scenario-specific crowd datasets, facilitate data-driven
research, and next build functional machine learning models. The main goal of this work was to develop an

Publication doi: 10.5220/0011691500003417

extension of crowd simulation (named CrowdSim2) and prove its usability in the application of people-tracking
algorithms. The simulator is developed using the very popular Unity 3D engine with particular emphasis on the
aspects of realism in the environment, weather conditions, traffic, and the movement and models of individual
agents. Finally, three methods of tracking were used to validate generated dataset: IOU-Tracker, Deep-Sort, and
Deep-TAMA.
Keywords Crowd simulation · realism enhancement · people and car simulation · people tracking · deep learning

1 Introduction • simulation placed in 3 different localisations (with 3

views) for people movement and 2 localisations for
Using real crowd datasets can produce effective and reliable cars,
learning models, useful in the following applications such as • introduction of 4 realistic weather conditions including
object tracking [1] [2], image segmentation [3] [4], visual object sun, fog, rain, snow, and different day time,
counting [5] [6] [7], individuals activity or violence recogni-
tion [8, 9], crowd anomaly detection and prediction and wider • many possible options for application including object
crowd management solutions monitor. However, acquiring real detection and tracking, action detection, and recogni-
crowd data faces several challenges, including the expensive tion.
installation of a sensory infrastructure, the data pre-processing
costs, and the lack of real datasets that cover particular crowd
scenarios. Consequently, simulation tools have been adopted for 2 Related works
generating synthetic datasets to overcome the challenges asso-
ciated with their real counterparts. Using simulation tools that Databases containing visual data are needed to develop the de-
can significantly reduce the time required to generate scenario- tection and tracking method, including cars and pedestrians.
specific crowd datasets, mimic observed crowds in a realistic For that purpose the annotations of traced objects (like a pedes-
environment, facilitate data-driven research, and build func- trian) which includes an approximate bounding box are nec-
tional machine learning models [10, 11] based on generated essary. The Mall dataset [13] was collected from a publicly
data. Simulation offers flexibility in adjusting the scenarios, and accessible webcam with ground truth consisting of annotating
generating and reproducing datasets with defined requirements. 60,000 pedestrians. NWPU [14] includes approximately 5,000
images and 2,133,375 annotated heads. JHU-CROWD++ [15]
The main motivation for that work was to implement more realis- is another crowd dataset captured in different scenarios, and geo-
tic crowd simulation with additional features that can be applied graphical locations, under weather conditions such as fog, haze,
in many modern artificial intelligence approaches (including the snow, and rain. JHU-CROWD++ provides head-level labeling
evaluation of people tracking algorithms). The proposed crowd which includes an approximate bounding box. The GTA5 Crowd
simulator has the following advantages: Counting (GCC) [16] is an example of a large-scale visual syn-
thetic dataset (15,212 images, 7,625,843 persons) generated
• realism enhancement by application of motion match- using the well-known video game GTA5 (Grand Theft Auto 5).
ing, people and car generation, AGORASET [17] is also a visual synthetic dataset for crowd
video analysis. For a comprehensive overview of databases and
• automatic assessment of ground truth and detection simulators, see the following review articles [18, 19, 20, 21]. In
(Figure 1) in the known format of MOT Challenge [22] the analysis of the statistical properties of real word datasets
[12], is available. Recent advancements in crowd simulation unravel

*correspondence: [email protected]
Preprint – Development of a Realistic Crowd Simulation Environment for Fine-grained Validation of People Tracking Methods
2

Figure 1: The visualization of tracking of multiple pedestrians in images generated by application of CrowdSim2.

Figure 2: Exemplary views from CrowdSim2: junction and park with moving pedestrians and cars along with examples of snow
and rain weather conditions that can be generated.

a wide range of functionalities for virtual agents, delivering a Deep Association Metric as a tracking-by-detection method.
highly-realistic, natural virtual crowds. Deep-Sort is an extension of the SORT algorithm [27] made to
integrate appearance information based on a deep appearance
In the following work, simulated data will be used for the eval-
descriptor. The Deep-TAMA method [28]– which stands for
uation of different tracking algorithms [23]. Here the concept
Deep Temporal Appearance Matching Association contrary to
of tracking by detection will be used, which means that detec-
Deep-Sort and IOU-tracker during one stage performs tracking
tion will be available along with simulated data, and afterward
together with evaluation of the results. Another group of SORT
tracking algorithms are applied for joining detections in tracks.
applications - Observation-Centric SORT [29] - is used for a
Additionally, instead of whole-person detection, facial recog-
multiple object tracker. OC-SORT was built to fix limitations in
nition could be applied [24]. The first considered algorithm
the Kalman filter and SORT algorithm. It is an online tracker and
IOU-tracker was presented in [25] and it doesn’t use any image
it has improved non-linear motion and robustness over occlu-
information, which allows it to run simpler tracking algorithms.
sion. For wide application, the framework MMtracking [30] was
Thanks to this non-image approach, it uses much less computing
established which is an open-source video perception toolbox
power than other trackers. The authors of the Deep-Sort method
by PyTorch.
[26] - presented Simple Online and Realtime Tracking with
Preprint – Development of a Realistic Crowd Simulation Environment for Fine-grained Validation of People Tracking Methods
3

Table 1: Information summarizing a number of folders, seconds,

and frames of data for different weather conditions.
Number of folders seconds frames
Sun 2899 86 970 2 174 250
Rain 1633 48 990 1 224 750
Fog 1653 49 590 1 239 750
Snow 1646 49 380 1 234 500

3 Crowd simulator
The proposed crowd simulator CrowdSim22 is the next ver- Figure 3: Agents movement concerning the concept of motion
sion of crowd simulator CrowdSim [31] especially for testing matching and system of features.
multi-object tracking algorithms but also for action and object
detection. It uses the microscopic (or ‘agent-based’) crowd sim-
ulation methods that model the behavior of each person, from
3.2 Photo-realistic Rendering and Traffic
which collective behavior can then emerge [32, 20]. The simula-
tor is developed using the very popular Unity 3D engine with The main element is physical volumetric light that responds
particular emphasis on the aspects of realism in the environment, adequately to dynamically changing surroundings. There is a
weather conditions, traffic, and the movement and models of dynamic volumetric fog in the simulation. Additionally, the
individual agents. The proposed system can be used to generate simulation also includes snow and rain based on particle effects
a sequence of random images (datasets) for use in tracking and (Figure 2). Finally, thanks to the use of the High Definition
object detection algorithms evaluation but also in the crowd, car Render Pipeline (HDRP) in Unity engine and physical cameras,
counting, and other crowd and traffic analysis tasks. The gen- it is possible to map the lens and matrix settings of the real
erated output data is in the format of the MOT challenge. The camera to create photo-realistic output images. For the global
most important components of CrowdSim2, to support realism agent movement on a macro scale, the already built-in NavMesh
when rendering the resulting image, are described below. was used. On the micro-scale at the beginning of the study
of human behavior, Emergent human behavior in Navigation
was created by the application of the unique system of features.
3.1 Agents Motion and Interactions Cars can park in randomly selected parking bays as visible in
Figure 4. At the crossing, they also choose a random direction.
The component necessary for producing lively and realistic vir- Traffic is based on a created system of nodes located on the
tual crowds is animating the characters, thus creating 3D moving roads, particularly crossing, sharp turns, and parking places.
agents. Data-driven approaches include methods utilizing mo-
tion capture data to use during skeleton-based animation of 3D
human models [33]. This approach requires many variations of
data to represent movements in different activities. To ensure
the universality of the system and to generate animations based
on real human motion data, a motion matching algorithm was
used [34]. Motion matching is an alternative animation system
without the need for a state machine with vectors given in Figure
3. Thanks to this, it is possible to perform different activities at
the agent level, including dancing or fights. In the future, it is
planned to use the learned motion matching algorithm [35] with
additional styling [36, 37]. Currently, we only have two styles
of movement male and female, which rely on separate motion
databases. The interactions are carried out based on interaction
zones. These zones are placed in the city and define the type
of interaction (for example dance, fight). When an agent enters Figure 4: The navigation system for car movement with parking
the zone, a set of conditions is checked, and, depending on the area and pedestrian stops.
situation, the agent is either ignored interaction, added to the
queue, or starts interaction (see Figure 6). If an agent is added
to the queue, there are not enough agents in the zone to start 4 Results and discussion
interacting. The agent moves normally while waiting in the
queue, and if he moves too far away from the zone before the 4.1 Collected Dataset
interaction begins, he is removed from the zone queue.
CrowdSim2 was used as the main tool for the generation of many
different simulated situations in 3 main places in the virtual city.
2
The dataset is freely available in the Zenodo Repository at https: All places were observed by 3 cameras directed at the same
//doi.org/10.5281/zenodo.7262220 position to get the general view from different angles. All videos
Preprint – Development of a Realistic Crowd Simulation Environment for Fine-grained Validation of People Tracking Methods
4

100
100.0
90 99.5

MOTA
iou-tracker
80 99.0
MOTA

Deep-TAMA
70 Deep-Sort
98.5
60
98.0
50 10 20 30 40 50 60 70 80 90 100110120130140150160
10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160
Density
(a) MOTA for varying crowd density. (b) Varying density for MOTA on IOU-tracker.
100 100
80
95
MOTA

MOTA
60
90 40
10 20 30 40 50 60 70 80 90 100110120130140150160
20 10 20 30 40 50 60 70 80 90 100110120130140150160
Density Density
(c) Varying density for MOTA on Deep-Sort. (d) Varying density for MOTA on Deep-TAMA.
Figure 5: Results of investigated tracking methods (IOU-tracker, Deep-Sort and Deep-TAMA) on varying crowd density data
(from 1 - 160 agents) on MOTA parameter.

ent weather conditions utilizing clear sunny weather (with just

resulting sun reflections), rain and snow with cloudy weather
influencing also background of the scene and fog day. Finally, 3
methods of tracking were used to test generated dataset: IOU-
Tracker [25], Deep-Sort [26] and Deep-TAMA [28]. Presented
methods were chosen under two conditions - first the availability
of open source code and second finite time of execution. All
methods were applied in MOT Challenge ranking. The given
evaluation was divided into two separate subsections - 1) vali-
dation of tracking methods in changing the number of people
in simulation and 2) verification of different weather conditions.
Methods were tested for the chosen parameters applied in MOT
Challenge: a) MOTA - Multiple-Object Tracking Accuracy and
Figure 6: Random animated interaction between agents gener- b) IDs - ID switches.
ated in CrowdSim2. Crowd simulation was first run on the different numbers of
people, which varied from 1 to 160. Thanks to that it was
possible to verify what is the influence of the number of people
were recorded in the resolution of 800x600 in 25 frames per in tracking results. The result of the comparison is presented in
second. Each situation lasted 30 seconds - 750 pictures were Figure 5 in the form of influence on the MOTA parameter and
recorded. The number of pedestrians varied from 1 to 160, and also the distribution of results for the set of people. In the second
situations were influenced by weather conditions such as sun, step, data were divided into weather conditions - sun, rain, fog,
rain, fog, and snow. The dataset was generated in the format of and snow. Here it is also possible to distinguish differences in
MOT Challenge. The number of generated videos, folders and the results of methods. The final score was presented in a few
frames is presented in Table 1. aspects - in the form of boxplots for each weather condition on
the MOTA parameter (Figure 7) and on different parameters in
4.2 Results Figure 8 and Table 2.
The obtained dataset was validated in two different scenarios:
1) by verification of the influence of the crowd density on the
accuracy of tracking methods and 2) by application of 4 differ-
Preprint – Development of a Realistic Crowd Simulation Environment for Fine-grained Validation of People Tracking Methods
5

100 100
80
80
60
60
40
40 20
iou-tracker Deep-TAMA Deep-Sort iou-tracker Deep-TAMA Deep-Sort
(a) Fog weather condition. (b) Rain weather condition.

100
100
80 80
60 60
40 40
20 iou-tracker Deep-TAMA Deep-Sort 20
iou-tracker Deep-TAMA Deep-Sort
(c) Snow weather condition. (d) Sun weather condition.
Figure 7: The influence of different weather conditions (fog, rain, snow, and sun) on the MOTA parameter for exemplary data from
CrowdSim2 presented in the form of boxplots.

Table 2: Mean and standard deviation for tracking method results case the best results are achieved by the IOU-tracker taking into
concerning exemplary evaluation parameters MOTA (that should consideration just the bounding box position. In fact to get the
be high) and Ids (that should be low). conclusion regarding validation both approaches can be used
and still conclusion of which method can give better tracking is
Method MOTA ↑ Ids ↓
possible. In all cases, in terms of MOTA, it can be observed that
Deep-Sort 96.20 ± 2.64 362.05 ± 436.04
a better method generates higher results. On the other hand for
IOU-tracker 99.74 ± 0.36 13.07 ± 159.64
IDs, that value should be lowered.
Deep-TAMA 74.58 ± 13.36 241.41 ± 349.84
Crowd simulations and direct connection to a graphical engine al-
low the generation of very accurate detections (bounding boxes)
for all visible pedestrians and on each frame. After validation
4.3 Discussion
performance, it could be noticed that generation of all detections
The generated data was used for two different validations of has some disadvantages. The first drawback lies in the number
tracking methods. On one side crowd density can be the first of detections and their size - that means that also very small
point of analysis. For the MOTA parameter, it can be observed pedestrians can be generated and put in the ground truth. On
which method can obtain better results. In all cases, the IOU- the other hand, in practice, obtaining detection for each frame is
tracker generates better results which rely mainly on the as- mostly not possible. Due to evaluation processing, any changes
sumption that automatically all detections are given at once. in that protocol were not submitted but in the future generated
Deep-Sort has slightly worse results because it takes also image data could be randomly disturbed in terms of the number of
context into consideration. Deep-TAMA fails in the case of detections. On the other hand on generated data, some other
simulated data due to the size of smaller detections. It has to be detection methods could be applied to produce not ideal detec-
mentioned that the number of people is not uniformly distributed tions.
- the generation was run in the specific number of pedestrians
but not always all detections were present and because of it for
160 pedestrians, not so many trails we could obtain. 5 Summary
In terms of weather conditions, the order of accuracy of methods In the presented work an extension of crowd simulation Crowd-
fits the crowd density analysis. The most challenging conditions Sim2 was introduced with many advanced features applied for
are given for snow. It may result from the influence of remaining the reality enhancement of generated results. To show the prac-
snow on the background. The same trend is visible and satisfied tical application of generated simulated data, tracking methods
for different parameters from the MOTA challenge. Still in that were run for evaluation purposes. Algorithms were tested con-
Preprint – Development of a Realistic Crowd Simulation Environment for Fine-grained Validation of People Tracking Methods
6

100
300
90 iou-tracker
MOTA

Deep-TAMA 200 iou-tracker

IDs
Deep-Sort Deep-TAMA
80 Deep-Sort
100
70
Sun Rain Snow Fog Sun Rain Snow Fog
(a) MOTA - Multiple-Object Tracking Accuracy. (b) IDs - ID switches.
Figure 8: The impact of weather conditions in the validation of tracking methods for different evaluation parameters for the dataset
from CrowdSim2. For MOTA the obtained results should be maximized and for IDs, the values should be lowered.

cerning the crowd density and weather conditions showing dif- 633–644. Springer International Publishing, 2022. doi:
ferences in final results and ordering accuracy of methods. The 10.1007/978-3-031-06430-2_53. URL https://doi.
obtained results confirmed that synthetic data from CrowdSim2 org/10.1007%2F978-3-031-06430-2_53.
can be used in the validation process for many scenarios without [2] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He,
the need for real data. Besides tracking algorithms it can be and Piotr Dollar. Focal loss for dense object detection. In
applied for object detection, action detection, and recognition, 2017 IEEE International Conference on Computer Vision
as a part of the testing procedure, and also in the training of (ICCV). IEEE, oct 2017. doi: 10.1109/iccv.2017.324.
machine learning algorithms. In the future, generated simulated
dataset can be enhanced by the used of post-processing methods [3] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae
to improve reality. Lee. YOLACT: Real-time instance segmentation. In 2019
IEEE/CVF International Conference on Computer Vision
(ICCV). IEEE, oct 2019. doi: 10.1109/iccv.2019.00925.
Acknowledgements [4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos,
This work was supported by: European Union funds awarded Kevin Murphy, and Alan L. Yuille. DeepLab: Semantic
to Blees Sp. z o.o. under grant POIR.01.01.01-00-0952/20-00 image segmentation with deep convolutional nets, atrous
“Development of a system for analysing vision data captured by convolution, and fully connected CRFs. IEEE Transactions
public transport vehicles interior monitoring, aimed at detect- on Pattern Analysis and Machine Intelligence, 40(4):834–
ing undesirable situations/behaviours and passenger counting 848, apr 2018. doi: 10.1109/tpami.2017.2699184.
(including their classification by age group) and the objects [5] Luca Ciampi, Claudio Gennaro, Fabio Carrara, Fabrizio
they carry”); EC H2020 project “AI4media: a Centre of Ex- Falchi, Claudio Vairo, and Giuseppe Amato. Multi-
cellence delivering next generation AI Research and Training camera vehicle counting using edge-AI. Expert Sys-
at the service of Media, Society and Democracy” under GA tems with Applications, 207:117929, nov 2022. doi:
951911; research project (RAU-6, 2020) and projects for young 10.1016/j.eswa.2022.117929. URL https://doi.org/
scientists of the Silesian University of Technology (Gliwice, 10.1016%2Fj.eswa.2022.117929.
Poland); research project INAROS (INtelligenza ARtificiale per
il mOnitoraggio e Supporto agli anziani), Tuscany POR FSE [6] Marco Avvenuti, Marco Bongiovanni, Luca Ciampi, Fab-
CUP B53D21008060008. Publication supported under the Ex- rizio Falchi, Claudio Gennaro, and Nicola Messina. A
cellence Initiative - Research University program implemented spatio- temporal attentive network for video-based crowd
at the Silesian University of Technology, year 2022. This re- counting. In 2022 IEEE Symposium on Computers and
search was supported by the European Union from the European Communications (ISCC). IEEE, jun 2022. doi: 10.1109/
Social Fund in the framework of the project "Silesian Univer- iscc55528.2022.9913019. URL https://doi.org/10.
sity of Technology as a Center of Modern Education based on 1109%2Fiscc55528.2022.9913019.
research and innovation” POWR.03.05.00- 00-Z098/17. We are [7] Luca Ciampi, Fabio Carrara, Valentino Totaro, Raffaele
thankful for students participating in design of Crowd Simulator: Mazziotti, Leonardo Lupori, Carlos Santiago, Giuseppe
P. Bartosz, S. Wróbel, M. Wola, A. Gluch and M. Matuszczyk. Amato, Tommaso Pizzorusso, and Claudio Gennaro.
Learning to count biological structures with raters’ un-
certainty. Medical Image Analysis, 80:102500, aug 2022.
References doi: 10.1016/j.media.2022.102500. URL https://doi.
[1] Donato Cafarelli, Luca Ciampi, Lucia Vadicamo, Clau- org/10.1016%2Fj.media.2022.102500.
dio Gennaro, Andrea Berton, Marco Paterni, Chiara Ben- [8] Luca Ciampi, Paweł Foszner, Nicola Messina, Michał
venuti, Mirko Passera, and Fabrizio Falchi. MOBDrone: Staniszewski, Claudio Gennaro, Fabrizio Falchi, Gian-
A drone video dataset for man OverBoard rescue. In luca Serao, Michał Cogiel, Dominik Golba, Agnieszka
Image Analysis and Processing – ICIAP 2022, pages Szczesna,
˛ and Giuseppe Amato. Bus violence: An open
Preprint – Development of a Realistic Crowd Simulation Environment for Fine-grained Validation of People Tracking Methods
7

benchmark for video violence detection on public transport. [20] W. van Toll and J. Pettré. Algorithms for microscopic
Sensors, 22(21):8345, oct 2022. doi: 10.3390/s22218345. crowd simulation: Advancements in the 2010s. Computer
URL https://doi.org/10.3390%2Fs22218345. Graphics Forum, 40(2):731–754, may 2021. doi: 10.1111/
[9] Paweł Foszner, Michał Staniszewski, Agnieszka Szczesna,
˛ cgf.142664. URL https://doi.org/10.1111%2Fcgf.
Michał Cogiel, Dominik Golba, Luca Ciampi, Nicola 142664.
Messina, Claudio Gennaro, Fabrizio Falchi, Giuseppe Am- [21] Shanwen Yang, Tianrui Li, Xun Gong, Bo Peng, and Jie
ato, and Gianluca Serao. Bus Violence: a large-scale Hu. A review on crowd simulation and modeling. Graphi-
benchmark for video violence detection in public transport, cal Models, 111:101081, 2020.
September 2022. URL https://doi.org/10.5281/ [22] Javad Amirian, Bingqing Zhang, Francisco Valente Cas-
zenodo.7044203. tro, Juan Jose Baldelomar, Jean-Bernard Hayet, and Julien
[10] Anish R Khadka, MM Oghaz, W Matta, M Cosentino, Pettré. Opentraj: Assessing prediction complexity in hu-
Paolo Remagnino, and Vasileios Argyriou. Learning how man trajectories datasets. In Proceedings of the Asian
to analyse crowd behaviour using synthetic data. In Pro- Conference on Computer Vision, 2020.
ceedings of the 32nd International Conference on Com- [23] Michał Staniszewski, Mateusz Kloszczyk, Jakub Segen,
puter Animation and Social Agents, pages 11–14, 2019. Kamil Wereszczyński, Aldona Drabik, and Marek Kul-
[11] Luca Ciampi, Nicola Messina, Fabrizio Falchi, Claudio backi. Recent developments in tracking objects in a video
Gennaro, and Giuseppe Amato. Virtual to real adaptation sequence. In Intelligent Information and Database Sys-
of pedestrian detectors. Sensors, 20(18):5250, sep 2020. tems, pages 427–436. Springer Berlin Heidelberg, 2016.
doi: 10.3390/s20185250. URL https://doi.org/10. doi: 10.1007/978-3-662-49390-8_42.
3390%2Fs20185250. [24] Damian Peszor,
˛ Michał Staniszewski, and Marzena Wo-
[12] P. Dendorfer, H. Rezatofighi, A. Milan, J. Shi, D. Cremers, jciechowska. Facial reconstruction on the basis of video
I. Reid, S. Roth, K. Schindler, and L. Leal-Taixé. Mot20: surveillance system for the purpose of suspect identifica-
A benchmark for multi object tracking in crowded scenes. tion. In Ngoc Thanh Nguyen, Bogdan Trawiński, Hamido
arXiv:2003.09003[cs], March 2020. URL http://arxiv. Fujita, and Tzung-Pei Hong, editors, Intelligent Informa-
org/abs/1906.04567. arXiv: 2003.09003. tion and Database Systems, pages 467–476, Berlin, Hei-
delberg, 2016. Springer Berlin Heidelberg. ISBN 978-3-
[13] Ke Chen, Chen Change Loy, Shaogang Gong, and Tony 662-49390-8.
Xiang. Feature mining for localised crowd counting. In
Procedings of the British Machine Vision Conference 2012. [25] Erik Bochinski, Tobias Senst, and Thomas Sikora.
British Machine Vision Association, 2012. doi: 10.5244/c. Extending iou based multi-object tracking by vi-
26.21. URL https://doi.org/10.5244%2Fc.26.21. sual information. In IEEE International Conference
on Advanced Video and Signals-based Surveillance,
[14] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu- pages 441–446, Auckland, New Zealand, November
crowd: A large-scale benchmark for crowd counting and 2018. URL http://elvera.nue.tu-berlin.de/
localization. IEEE transactions on pattern analysis and files/1547Bochinski2018.pdf.
machine intelligence, 43(6):2141–2149, 2020. [26] Nicolai Wojke and Alex Bewley. Deep cosine metric learn-
[15] Vishwanath Sindagi, Rajeev Yasarla, and Vishal MM Pa- ing for person re-identification. In 2018 IEEE Winter Con-
tel. Jhu-crowd++: Large-scale crowd counting dataset ference on Applications of Computer Vision (WACV), pages
and a benchmark method. IEEE Transactions on Pattern 748–756. IEEE, 2018. doi: 10.1109/WACV.2018.00087.
Analysis and Machine Intelligence, 2020. [27] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Sim-
[16] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning ple online and realtime tracking with a deep association
from synthetic data for crowd counting in the wild. In metric. In 2017 IEEE International Conference on Image
Proceedings of the IEEE/CVF conference on computer Processing (ICIP), pages 3645–3649. IEEE, 2017. doi:
vision and pattern recognition, pages 8198–8207, 2019. 10.1109/ICIP.2017.8296962.
[17] Nicolas Courty, Pierre Allain, Clement Creusot, and [28] Young-Chul Yoon, Du Yong Kim, Kwangjin Yoon, Young-
Thomas Corpetti. Using the agoraset dataset: Assessing min Song, and Moongu Jeon. Online multiple pedes-
for the quality of crowd video analysis methods. Pattern trian tracking using deep temporal appearance match-
Recognition Letters, 44:161–170, 2014. ing association. CoRR, abs/1907.00831, 2019. URL
http://arxiv.org/abs/1907.00831.
[18] Amna Bamaqa, Mohamed Sedky, Tomasz Bosakowski,
Benhur Bakhtiari Bastaki, and Nasser O Alshammari. [29] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and
Simcd: Simulated crowd data for anomaly detection and Ben Upcroft. Simple online and realtime tracking. In
prediction. Expert Systems with Applications, 203:117475, 2016 IEEE International Conference on Image Processing
2022. (ICIP), pages 3464–3468, 2016. doi: 10.1109/ICIP.2016.
7533003.
[19] Marilena Lemonari, Rafael Blanco, Panayiotis Charalam-
bous, Nuria Pelechano, Marios Avraamides, Julien Pettré, [30] MMTracking Contributors. MMTracking: OpenMM-
and Yiorgos Chrysanthou. Authoring virtual crowds: A Lab video perception toolbox and benchmark. https:
survey. Computer Graphics Forum, 41(2):677–701, may //github.com/open-mmlab/mmtracking, 2020.
2022. doi: 10.1111/cgf.14506. URL https://doi.org/ [31] Michał Staniszewski, Paweł Foszner, Karol Kostorz, Ag-
10.1111%2Fcgf.14506. nieszka Michalczuk, Kamil Wereszczyński, Michał Cogiel,
Preprint – Development of a Realistic Crowd Simulation Environment for Fine-grained Validation of People Tracking Methods
8

Dominik Golba, Konrad Wojciechowski, and Andrzej Po-

lański. Application of crowd simulations in the evaluation
of tracking algorithms. Sensors, 20(17):4960, 2020.
[32] Raza Abdulla Saeed, Diego Reforgiato Recupero, and
Paolo Remagnino. Simulating crowd behaviour combin-
ing both microscopic and macroscopic rules. Information
Sciences, 583:137–158, 2022.
[33] Kamil Wereszczyński, Agnieszka Michalczuk, Paweł
Foszner, Dominik Golba, Michał Cogiel, and Michał
Staniszewski. Elsa: Euler-lagrange skeletal animations
- novel and fast motion model applicable to vr/ar devices.
In Computational Science – ICCS 2021, pages 120–133,
Cham, 2021. Springer International Publishing. ISBN
978-3-030-77977-1.
[34] Simon Clavet. Motion matching and the road to next-gen
animation. In Proc. of GDC, volume 2016, 2016.
[35] Daniel Holden, Oussama Kanoun, Maksym Perepichka,
and Tiberiu Popa. Learned motion matching. ACM Trans-
actions on Graphics (TOG), 39(4):53–1, 2020.
[36] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-
Or, and Baoquan Chen. Unpaired motion style transfer
from video to animation. ACM Transactions on Graphics
(TOG), 39(4):64–1, 2020.
[37] Daniel Holden, Ikhsanul Habibie, Ikuo Kusajima, and Taku
Komura. Fast neural style transfer for motion data. IEEE
computer graphics and applications, 37(4):42–49, 2017.