Multi-fish Detection And Tracking Using Track-mask Region Convolutional Neural Network
Deep learning has become more common in recent years due to its excellent results in many areas. This thesis primarily focuses on multi-fish detection and tracking methods in underwater videos. The existing multi-fish detection methods for underwater videos have a low detection rate and consumes...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | http://eprints.usm.my/60661/1/NAWAF%20FARHAN%20FANKUR%20ALSHDAIFAT%20-%20TESIS24.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Deep learning has become more common in recent years due to its excellent results
in many areas. This thesis primarily focuses on multi-fish detection and tracking methods
in underwater videos. The existing multi-fish detection methods for underwater
videos have a low detection rate and consumes time in the training and testing process
due to the underwater conditions and the overfitting during training. Many multi-fish
detection and tracking methods for underwater videos (based on deep learning) where
low accuracy for multi-fish tracking and occlusion instances during multi-fish tracking
leads to inability to distinguish edges, and inability to handle each detected object over
time. Therefore, this research aims to improve and enhance methods for multi-fish
detection and tracking in underwater videos based on the latest deep learning algorithms.
The proposed improved multi-fish detection method involves three main steps:
1) Improving ResNet-101 backbone for better fish detection, 2) Enhancing the Region
Proposal Network (RPN) method based on Faster R-CNN for multi-fish detection and
3) An improved multi-fish detection method in terms of accuracy and with a lower
training and testing times by utilising the aforementioned methods. The proposed
multi-fish tracking method (Track-Mask R-CNN) also exhibits similar enhanced characteristics
compared to the state-of-art methods (using fish dataset). An accuracy of
86.7% and 78.9% have been achieved for the proposed multi-fish detection and tracking
respectively. |
|---|