Night-time Vehicle Detection Based on Observable Light Cues Using Deep Learning
(2023)Department of Automatic Control
- Abstract
- This thesis investigates the issue of computer vision-based detection of oncoming cars during night-time, a critical road safety issue for automated high-beam assist. We propose a holistic image classification approach that uses deep learning methods to detect light artifacts from an oncoming car’s headlights before the car is entirely visible. We explore six different model architectures, including both convolutional neural networks and transformer-based models. We train them using transfer learning with both public and internal datasets using models pre-trained on ImageNet. We evaluate the generalization ability of the models and find that they can achieve up to 71% accuracy when trained on the public dataset and evaluated on the... (More)
- This thesis investigates the issue of computer vision-based detection of oncoming cars during night-time, a critical road safety issue for automated high-beam assist. We propose a holistic image classification approach that uses deep learning methods to detect light artifacts from an oncoming car’s headlights before the car is entirely visible. We explore six different model architectures, including both convolutional neural networks and transformer-based models. We train them using transfer learning with both public and internal datasets using models pre-trained on ImageNet. We evaluate the generalization ability of the models and find that they can achieve up to 71% accuracy when trained on the public dataset and evaluated on the class-balanced internal dataset. Our results show that both convolution-based and transformer-based models have potential in performance for this task, with the best models reaching up to 88% accuracy when trained with the full public dataset and evaluated with the class-balanced public test set. Our research contributes to the field by introducing an approach to detection of oncoming cars and comparing different model architectures for this task. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9136434
- author
- Ivarsson, Celine and Zacke, Jennifer
- supervisor
- organization
- year
- 2023
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6195
- other publication id
- 0280-5316
- language
- English
- id
- 9136434
- date added to LUP
- 2023-09-06 14:16:26
- date last changed
- 2023-09-06 14:16:26
@misc{9136434, abstract = {{This thesis investigates the issue of computer vision-based detection of oncoming cars during night-time, a critical road safety issue for automated high-beam assist. We propose a holistic image classification approach that uses deep learning methods to detect light artifacts from an oncoming car’s headlights before the car is entirely visible. We explore six different model architectures, including both convolutional neural networks and transformer-based models. We train them using transfer learning with both public and internal datasets using models pre-trained on ImageNet. We evaluate the generalization ability of the models and find that they can achieve up to 71% accuracy when trained on the public dataset and evaluated on the class-balanced internal dataset. Our results show that both convolution-based and transformer-based models have potential in performance for this task, with the best models reaching up to 88% accuracy when trained with the full public dataset and evaluated with the class-balanced public test set. Our research contributes to the field by introducing an approach to detection of oncoming cars and comparing different model architectures for this task.}}, author = {{Ivarsson, Celine and Zacke, Jennifer}}, language = {{eng}}, note = {{Student Paper}}, title = {{Night-time Vehicle Detection Based on Observable Light Cues Using Deep Learning}}, year = {{2023}}, }