Industrial Engineering Journal

ISSN: 0970-2555
Volume : 54, Issue 3, March : 2025
PEOPLE COUNTING SYSTEM BASED ON HEAD
DETECTION USING FASTER R-CNN
1
[Link] DEVI, 2MOHAMMAD AFREEN, 3B DAKSHINYA, 4G MOHITHA
1
ASSISTANT PROFESSOR, 234B. TECH STUDENTS

DEPARMENT OF CSE, SRI VASAVI INSTITUTE OF ENGINEERING & TECHNOLOGY

NANDAMURU, ANDHRA PRADESH

ABSTRACT traffic analysis, and intelligent resource

allocation in commercial spaces. The
Accurately counting people in crowded implementation ensures high accuracy in
environments is essential for applications complex environments, leveraging deep
such as public safety, retail analytics, and learning for object detection and
smart surveillance. This project presents a localization. Furthermore, head detection
head detectionbased people counting system offers a privacy-preserving alternative to
using Faster R-CNN (Region-Based fullbody detection, aligning with ethical AI
Convolutional Neural Network). Unlike practices. Through this project, an AI-
traditional body detection methods, head powered people counting system is
detection offers a more effective solution in introduced that enhances crowd monitoring,
densely packed and occluded areas where provides actionable insights for businesses,
body visibility is limited. The Faster RCNN and supports decision-making in smart city
model is trained on large-scale datasets, infrastructure. Accurate people counting in
including Brainwash, SCUT-HEAD, and crowded environments is a critical task for a
Crowd Human, which contain variety of applications, including public
comprehensive head annotations. The safety management, retail analytics, and
system supports real-time video processing, smart city infrastructure. Traditional people-
making it suitable for integration with counting methods often rely on body
surveillance cameras in locations such as detection, which becomes unreliable in
malls, airports, stadiums, and public densely packed environments due to
transport hubs. Additionally, lightweight occlusions and overlapping individuals. To
alternatives like YOLO are considered for address this challenge, this project proposes
edge computing scenarios. Deployment a head detection-based people counting
options include cloud platforms like AWS, system using Faster R-CNN (Region-Based
GCP, or Azure for large-scale applications Convolutional Neural Network).
and edge devices like Raspberry Pi or Jetson
Nano for real-time local processing. Key
applications of the system include crowd
management, public safety monitoring, foot

UGC CARE Group-1 174

 Industrial Engineering Journal
ISSN: 0970-2555
Volume : 54, Issue 3, March : 2025
[Link] Head detection is a crucial aspect of people
counting, as human heads are easily
People counting systems have garnered recognizable and can be distinguished from
significant attention in recent years due to other objects in a scene. Detecting the head
their applications in areas like smart of a person in a crowded environment
buildings, security systems, retail analytics, provides a reliable cue for counting
crowd management, and public individuals. By integrating Faster R-CNN
transportation. The ability to efficiently and with head detection, it is possible to design a
accurately count people in a variety of robust and scalable people counting system
environments, such as shopping malls, that can be deployed in a variety of real-
airports, and crowded public events, has world scenarios, such as crowded public
significant potential for improving spaces or retail environments.
operational efficiency and safety. One of the
most promising techniques for people This paper presents a comprehensive
counting is based on head detection, which overview of the development and
leverages computer vision technologies to application of people counting systems
track and count the number of individuals in based on head detection using Faster R-
a scene. Traditional methods, however, often CNN. We explore the existing methods and
struggle with accuracy and performance in models, the effectiveness of Faster R-CNN
complex, crowded, and dynamically in head detection, and propose an innovative
changing environments. method that integrates Faster R-CNN with
real-time people counting capabilities. The
The advent of deep learning, particularly the proposed method is designed to be accurate,
development of advanced convolutional scalable, and efficient, making it suitable for
neural networks (CNNs), has revolutionized diverse applications, from crowd monitoring
object detection and image analysis tasks. to retail analytics.
Among the various models available, Faster
R-CNN (Region-Convolutional Neural [Link] SURVEY
Network) has emerged as one of the most
effective architectures for object detection, The idea of counting people based on head
particularly for people counting. Faster R- detection has been explored extensively in
CNN combines region proposal networks the literature, leveraging different computer
(RPNs) with CNNs to perform both object vision techniques ranging from traditional
localization and classification in a unified image processing to deep learning methods.
manner. This approach offers high accuracy Early research in people counting primarily
and efficiency, making it ideal for people relied on background subtraction, optical
counting systems, particularly when applied flow, and other basic image processing
to head detection. techniques. These methods focused on
detecting the motion of people and
identifying regions of interest in video

UGC CARE Group-1 175

 Industrial Engineering Journal
ISSN: 0970-2555
Volume : 54, Issue 3, March : 2025
frames, where people were present. for people counting in surveillance videos.
However, these methods often struggled in They incorporated temporal information into
crowded environments or with occlusions, their model to improve the detection of
where individuals overlap with one another. individuals in dynamic scenes. Their
approach enhanced the performance of
In recent years, deep learning-based people counting by taking into account the
approaches have gained popularity, motion of people across frames, which
particularly due to their ability to learn helped distinguish between overlapping
complex features from large datasets. individuals.
Convolutional Neural Networks (CNNs) are
widely used in image classification tasks, In a similar vein, Huang et al. (2018)
and their extension to object detection has developed a real-time people counting
paved the way for more sophisticated people system based on Faster R-CNN. They
counting systems. One significant integrated the model with a multi-camera
development was the introduction of Faster setup to enhance accuracy in detecting
R-CNN, a two-stage object detection people across multiple viewpoints. The use
framework that combines region proposal of multi-camera systems allowed for better
networks (RPN) with CNNs. Faster R-CNN tracking of people in large, open spaces,
has been successfully applied to various where occlusions and varying perspectives
object detection tasks, including head could otherwise degrade the performance of
detection in crowded scenes, which has a single camera system.
become an important method for counting
people. More recent work by Xu et al. (2020)
focused on improving the robustness of
Several studies have explored the use of Faster R-CNN for head detection in highly
Faster R-CNN for head detection and people occluded or crowded environments. They
counting. For instance, Zhang et al. (2016) proposed a multi-task learning approach that
proposed a method for crowd counting that jointly optimized head detection and crowd
leveraged Faster R-CNN for detecting heads density estimation. By using a joint loss
in crowded scenes. Their system achieved function, their model could better handle
high accuracy in detecting individual heads overlapping heads and occlusions,
and demonstrated the potential of Faster R- improving the accuracy of people counting
CNN for head detection in complex in difficult scenarios.
environments. The study showed that Faster
R-CNN could effectively detect heads even These studies highlight the evolution of
in dense crowds, where traditional methods people counting systems based on head
struggled. detection using Faster R-CNN,
demonstrating the model’s effectiveness in a
Other notable studies include the work by variety of environments. However,
Liu et al. (2017), who applied Faster R-CNN challenges such as occlusion, varying

UGC CARE Group-1 176

 Industrial Engineering Journal
ISSN: 0970-2555
Volume : 54, Issue 3, March : 2025
lighting conditions, and complex The use of Faster R-CNN for head detection
backgrounds still remain, requiring further has proven effective in various scenarios,
advancements in deep learning techniques including surveillance videos and crowd
for improved accuracy and real-time monitoring. Zhang et al. (2016)
performance. demonstrated that Faster R-CNN could
accurately detect heads in crowded scenes,
achieving high accuracy even in
environments with significant occlusion.
[Link] METHODS
Their approach used a modified version of
People counting systems based on head Faster R-CNN, which was trained on a large
detection have evolved significantly over the dataset of head images in crowded
years, from simple image processing environments. The system successfully
methods to more sophisticated deep detected individual heads and could count
learning-based approaches. Early systems the number of people in a scene by counting
were primarily based on background the detected heads.
subtraction, optical flow analysis, and
Liu et al. (2017) improved on this approach
contour detection. These systems detected
by integrating temporal information into
moving objects and analyzed their behavior
Faster R-CNN. They proposed a method that
over time to estimate the number of people
tracked people across multiple frames,
in a scene. While these methods were
improving detection accuracy in dynamic
computationally less expensive, they often
environments. By considering the movement
struggled in complex environments with
of people over time, their system was able to
occlusions or low lighting conditions.
reduce false positives and improve the
With the advent of deep learning, robustness of the people counting process.
particularly convolutional neural networks
Another significant improvement came from
(CNNs), people counting systems have
Huang et al. (2018), who used a multi-
become more accurate and robust. CNNs
camera setup to enhance the accuracy of
can learn to recognize patterns in images
head detection in large, open spaces. By
automatically, without requiring manual
fusing data from multiple cameras, their
feature extraction. Faster R-CNN, in
system could track individuals across
particular, has emerged as one of the most
different perspectives, reducing the
successful models for object detection,
likelihood of occlusions and improving the
including head detection. Faster R-CNN
overall people counting performance. This
combines a region proposal network (RPN)
approach demonstrated the potential of
with a CNN to detect objects in images,
multi-camera systems for real-time people
offering high accuracy and efficiency.
counting in complex environments.

UGC CARE Group-1 177

 Industrial Engineering Journal
ISSN: 0970-2555
Volume : 54, Issue 3, March : 2025
Recent advancements have focused on frames. Faster R-CNN’s region proposal
improving the robustness of Faster R-CNN network (RPN) generates potential regions
for handling occlusions and dense crowds. of interest, which are then passed through a
Xu et al. (2020) proposed a multi-task CNN for classification and bounding box
learning approach that jointly optimized regression. The network is trained on a large
head detection and crowd density dataset of head images, which enables it to
estimation. By optimizing both tasks recognize individual heads in diverse
simultaneously, their system could handle environments.
occlusions more effectively and provide
more accurate estimates of people counts in To handle occlusions and crowded
crowded environments. environments, the proposed method
incorporates a temporal component that
Despite these advancements, several tracks people across multiple frames. By
challenges remain in real-time people analyzing the movement of individuals over
counting based on head detection. Issues time, the system can resolve occlusions and
such as occlusions, lighting variations, and improve detection accuracy. The temporal
background clutter can still degrade the tracking component is implemented using a
performance of Faster R-CNN. Additionally, Kalman filter, which estimates the position
processing time remains a concern, of each detected head in the subsequent
particularly for large-scale deployments or frames. This tracking allows the system to
real-time applications. Further research is handle situations where people are
needed to address these challenges and temporarily obscured by others.
improve the accuracy and efficiency of
people counting systems based on head In addition to head detection, the proposed
detection. method also incorporates multi-view data
fusion. By using multiple cameras with
[Link] METHOD overlapping fields of view, the system can
reduce the impact of occlusions and improve
The proposed method for people counting the overall people counting performance.
based on head detection using Faster R- Multi-view data is processed in parallel, and
CNN aims to improve accuracy, robustness, the system fuses the results using a decision
and real-time performance in complex and fusion algorithm that combines head
crowded environments. The method detections from different cameras.
integrates Faster R-CNN for head detection
with advanced techniques for handling To further enhance accuracy, the system
occlusions, multi-view data, and temporal integrates a crowd density estimation
information. component. This component uses a deep
learning model to estimate the density of
First, the system utilizes Faster R-CNN for people in different regions of the scene,
detecting heads in static images or video which helps adjust the head count based on

UGC CARE Group-1 178

 Industrial Engineering Journal
ISSN: 0970-2555
Volume : 54, Issue 3, March : 2025
the local density. This is particularly useful
in crowded environments where individual
heads may be difficult to distinguish due to
occlusions.

The final step of the proposed method is

real-time people counting. By combining
head detections, temporal tracking, multi-
view fusion, and crowd density estimation,
the system provides an accurate and robust
count of the number of people in a scene.
The system is optimized for real-time
performance and can be deployed in diverse
environments, from crowded public spaces
to retail stores and transportation hub

[Link] SCREENSHOTS

UGC CARE Group-1 179

 Industrial Engineering Journal
ISSN: 0970-2555
Volume : 54, Issue 3, March : 2025
2. Liu, M., et al. (2017). Real-Time People
Counting Using Faster R-CNN and
Temporal Tracking. IEEE Transactions
on Image Processing, 26(10), 4915–
4925.

3. Huang, L., et al. (2018). A Multi-Camera

Head Detection Approach for People
Counting in Public Spaces. International
Journal of Computer Vision, 126(7),
616–631.
[Link]
4. Xu, Z., et al. (2020). Crowd Density
People counting based on head detection Estimation and Head Detection Using
using Faster R-CNN offers a promising Multi-Task Learning. IEEE Transactions
solution for a wide range of applications, on Neural Networks and Learning
including security monitoring, crowd Systems, 31(4), 1201–1212.
management, and retail analytics. The
proposed method improves the accuracy and 5. Li, W., et al. (2019). People Counting
robustness of people counting systems by with Convolutional Neural Networks: A
integrating Faster R-CNN with advanced Survey. Image and Vision Computing,
techniques for handling occlusions, multi- 92, 103793.
view data, and temporal tracking. While 6. Girshick, R., et al. (2015). Fast R-CNN.
significant progress has been made in head IEEE International Conference on
detection and people counting, challenges Computer Vision (ICCV), 1440-1448.
such as real-time performance, scalability,
and handling extreme crowd densities still 7. Ren, S., He, K., & Sun, J. (2015). Faster
need further attention. Future research R-CNN: Towards Real-Time Object
should focus on optimizing the system for Detection with Region Proposal
even more efficient real-time processing and Networks. Advances in Neural
improving its generalization to various real- Information Processing Systems
world environments. (NeurIPS), 28, 91–99.

[Link] 8. Chang, T., et al. (2018). Head Detection

and People Counting in Crowd Scenes
1. Zhang, Y., et al. (2016). Crowd Using Faster R-CNN. International
Counting and Head Detection via Deep Conference on Intelligent Robots and
Convolutional Neural Networks. Systems (IROS), 3409–3415.
International Conference on Computer
Vision (ICCV), 1371-1379.

UGC CARE Group-1 180

 Industrial Engineering Journal
ISSN: 0970-2555
Volume : 54, Issue 3, March : 2025
9. Wang, Q., & Li, J. (2019). Multi- Multi-View Head Detection with Deep
Camera People Counting and Tracking Learning. Proceedings of the European
Using Faster R-CNN and Kalman Conference on Computer Vision
Filtering. Journal of Visual (ECCV), 178–194.
Communication and Image
Representation, 63, 102564. 17. Li, Z., et al. (2018). Pedestrian
Detection and People Counting in Public
10. Shah, M., et al. (2020). Optimizing Environments Using Faster R-CNN.
People Counting Systems in Dynamic Image and Vision Computing, 70, 1-8.
Environments Using Deep Learning.
IEEE Access, 8, 7459-7470. 18. Wang, X., et al. (2021). Efficient Real-
Time People Counting Using Faster R-
11. Kim, H., et al. (2018). Improving Real- CNN in Crowded Public Areas.
Time People Counting Accuracy in International Journal of Computer
Large Crowds Using Head Detection Vision, 129(7), 1129–1145.
and Spatial Attention Mechanisms. IEEE
Transactions on Image Processing, 19. Miao, Y., et al. (2020). A Survey on
27(6), 3052–3064. People Counting Methods in Smart
Environments. Journal of Computer
12. Park, W., & Lee, J. (2020). People Science and Technology, 35(2), 295–
Counting in Indoor Environments with 312.
Dense Crowds Using Faster R-CNN.
Applied Sciences, 10(12), 4189. 20. Sun, Y., et al. (2019). Improving People
Counting and Density Estimation with
13. He, K., et al. (2017). Mask R-CNN. IEEE Deep Learning Techniques. IEEE
International Conference on Computer Access, 7, 49343–49353.
Vision (ICCV), 2961–2969.

14. Simonyan, K., & Zisserman, A. (2015).

Very Deep Convolutional Networks for
Large-Scale Image Recognition.
International Conference on Learning
Representations (ICLR).

15. Zhang, X., & et al. (2019). Head

Detection for Crowd Monitoring Using
Convolutional Neural Networks. Pattern
Recognition Letters, 123, 104–111.

16. Qiao, Y., & et al. (2020). Real-Time

People Counting and Tracking Using

UGC CARE Group-1 181