Deep learning for unmanned aerial vehicle-based object counting
The application of object counting to various fields, such as traffic flow monitoring, crowd density estimation, and product counting, has attracted considerable attention. The proposed technology has the potential to enhance productivity by facilitating fully automated object counting and reduci...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Master by Coursework |
| Language: | English |
| Published: |
Nanyang Technological University
2023
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/169217 |
| _version_ | 1811687330234236928 |
|---|---|
| author | Yuan, Ruizhi |
| author2 | Yap Kim Hui |
| author_facet | Yap Kim Hui Yuan, Ruizhi |
| author_sort | Yuan, Ruizhi |
| collection | NTU |
| description | The application of object counting to various fields, such as traffic flow monitoring,
crowd density estimation, and product counting, has attracted considerable attention. The proposed technology has the potential to enhance productivity by facilitating fully automated object counting and reducing the risk of human error. The field of object counting based on Unmanned Aerial Vehicles (UAVs) is gaining momentum due to its numerous advantages. For example, UAVs provide a high-altitude viewpoint, which allows for a wider field of view. Moreover, UAVs can be quickly deployed in complex environments like forests, mountains, and seas.
The use of deep learning is powerful for achieving complex tasks, such as object counting. Therefore, this dissertation aims to develop deep learning methods for UAV-based object counting. The dissertation employs the Heatmap Learner Convolutional
Neural Network (HLCNN), primarily training the fine-tuned VGG16 model to generate density maps based on CARPK and PUCPR+ datasets. The density maps generate a peak map that identifies the number of objects in a picture.
Due to different weather and illumination, the UAV-based object counting task will encounter inaccurate results. Therefore, the study proposes two improved methods for
addressing this issue. Firstly, introducing residual blocks can increase the depth of the
network and prevent gradient explosion and gradient vanishing. The second is the implementation of dilated convolutions, which offer a larger receptive field. The experimental findings indicate that introducing dilated convolution after the tenth layer can enhance the model's performance.
Keywords: Object Counting, VGG16, Residual Block, Dilated Convolution, UAV |
| first_indexed | 2024-10-01T05:14:36Z |
| format | Thesis-Master by Coursework |
| id | ntu-10356/169217 |
| institution | Nanyang Technological University |
| language | English |
| last_indexed | 2024-10-01T05:14:36Z |
| publishDate | 2023 |
| publisher | Nanyang Technological University |
| record_format | dspace |
| spelling | ntu-10356/1692172023-07-14T15:43:00Z Deep learning for unmanned aerial vehicle-based object counting Yuan, Ruizhi Yap Kim Hui School of Electrical and Electronic Engineering EKHYap@ntu.edu.sg Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision The application of object counting to various fields, such as traffic flow monitoring, crowd density estimation, and product counting, has attracted considerable attention. The proposed technology has the potential to enhance productivity by facilitating fully automated object counting and reducing the risk of human error. The field of object counting based on Unmanned Aerial Vehicles (UAVs) is gaining momentum due to its numerous advantages. For example, UAVs provide a high-altitude viewpoint, which allows for a wider field of view. Moreover, UAVs can be quickly deployed in complex environments like forests, mountains, and seas. The use of deep learning is powerful for achieving complex tasks, such as object counting. Therefore, this dissertation aims to develop deep learning methods for UAV-based object counting. The dissertation employs the Heatmap Learner Convolutional Neural Network (HLCNN), primarily training the fine-tuned VGG16 model to generate density maps based on CARPK and PUCPR+ datasets. The density maps generate a peak map that identifies the number of objects in a picture. Due to different weather and illumination, the UAV-based object counting task will encounter inaccurate results. Therefore, the study proposes two improved methods for addressing this issue. Firstly, introducing residual blocks can increase the depth of the network and prevent gradient explosion and gradient vanishing. The second is the implementation of dilated convolutions, which offer a larger receptive field. The experimental findings indicate that introducing dilated convolution after the tenth layer can enhance the model's performance. Keywords: Object Counting, VGG16, Residual Block, Dilated Convolution, UAV Master of Science (Communications Engineering) 2023-07-10T00:27:11Z 2023-07-10T00:27:11Z 2023 Thesis-Master by Coursework Yuan, R. (2023). Deep learning for unmanned aerial vehicle-based object counting. Master's thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/169217 https://hdl.handle.net/10356/169217 en application/pdf Nanyang Technological University |
| spellingShingle | Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision Yuan, Ruizhi Deep learning for unmanned aerial vehicle-based object counting |
| title | Deep learning for unmanned aerial vehicle-based object counting |
| title_full | Deep learning for unmanned aerial vehicle-based object counting |
| title_fullStr | Deep learning for unmanned aerial vehicle-based object counting |
| title_full_unstemmed | Deep learning for unmanned aerial vehicle-based object counting |
| title_short | Deep learning for unmanned aerial vehicle-based object counting |
| title_sort | deep learning for unmanned aerial vehicle based object counting |
| topic | Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision |
| url | https://hdl.handle.net/10356/169217 |
| work_keys_str_mv | AT yuanruizhi deeplearningforunmannedaerialvehiclebasedobjectcounting |