THE 3 rd INTERNATIONAL SCIENTIFIC CONFERENCES OF STUDENTS AND YOUNG RESEARCHERS dedicated to the 99
th
anniversary of the National Leader of Azerbaijan Heydar Aliyev
165
proposed as an object identification strategy that reduced computation
complexity by regarding detection as a single regression problem.
When it comes to deep learning-based object recognition, the YOLO
model is one of the most utilized, with significant speed advantages
that are ideal for real-time applications. The YOLO model was
implemented to detect pedestrians in this paper.
Transfer learning strategy is based on leveraging pre-trained
models on the Microsoft COCO dataset, followed by fine-tuning and
optimization of our YOLO-based model. It's the concept of breaking
free from the isolated learning paradigm and applying what you've
learned to address related problems. An overhead dataset is used to
train the model by using ready feature vector. As a result, the model
takes advantage of both pre-trained and newly taught information, and
both detection results are improved and delivered faster in terms of
model accuracy and response time, respectively.
The experimental findings show that the framework easily
recognizes persons strolling too near to each other and violating social
distance; moreover, the transfer learning approach improves the
detection model's overall efficiency and accuracy. The model achieves
detection accuracy of 92% without transfer learning and 95% with
transfer learning for a pre-trained model. The model is then deployed
with the help of Streamlit tool as a web application which has the
functionality of utilizing model on images, videos, and real-time web
cams. In the future, the work might be enhanced for diverse indoor and
outdoor conditions to improve the accuracy of the model and make its
usage available for all environments.
References [1] Ghiasi, G.; Lin, T.Y.; Le, Q.V. Nas: Learning scalable feature pyramid architecture for
object detection. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, Seoul, Korea, 27 October–2 November 2019; pp. 7036–7045.
[2] Ahmed, I., Ahmad, M., Rodrigues, J. J. P. C., Jeon, G., & Din, S. (2020). A deep learning-
based social distance monitoring framework for COVID-19. Sustainable Cities and
Society, 102571
[3] Zhou, B.; Wang, X.; Tang, X. Understanding collective crowd behaviors: Learning
a mixture model of dynamic pedestrian-agents