Object detection in thermal imagery for crowd density estimation
Timofeeva, Polina (2020)
Timofeeva, Polina
2020
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-202005128485
https://urn.fi/URN:NBN:fi:amk-202005128485
Tiivistelmä
This work represents a research on deep-learning based approaches for object detection for the purpose of crowd density estimation. Specifically, thermal imagery which allows to preserve personal identity was used to train a deep learning object detection system. The ultimate objective of this work is to provide the client company with the tool which would facilitate better understanding of how various office, meeting, common and work spaces are being used across the campus, optimize crowd flows and drive down maintenance costs.
Theoretical background related to both, traditional computer vision and novel deep-learning methods for object detection was studied and outlined in this work. Based on the outcomes of such comparison, the most promising method was implemented into production.
Extensive empirical evidence obtained through extensive testing of the proposed solution demonstrated that the model exhibits high accuracy, great generalization capabilities and robustness against various perturbations in input images. It was concluded that provided solution satisfies both, accuracy and inference time requirements and therefore qualified to be deployed into production.
Finally, possible directions for further development were outlined. Improved performance can be achieved by alternating backbone network architecture and expanding the training data set.
Theoretical background related to both, traditional computer vision and novel deep-learning methods for object detection was studied and outlined in this work. Based on the outcomes of such comparison, the most promising method was implemented into production.
Extensive empirical evidence obtained through extensive testing of the proposed solution demonstrated that the model exhibits high accuracy, great generalization capabilities and robustness against various perturbations in input images. It was concluded that provided solution satisfies both, accuracy and inference time requirements and therefore qualified to be deployed into production.
Finally, possible directions for further development were outlined. Improved performance can be achieved by alternating backbone network architecture and expanding the training data set.