Please use this identifier to cite or link to this item:
|Title:||Deep learning frameworks for lane detection in rainy images||Authors:||Tan, Herman Kai Liang||Keywords:||Engineering::Electrical and electronic engineering||Issue Date:||2021||Publisher:||Nanyang Technological University||Source:||Tan, H. K. L. (2021). Deep learning frameworks for lane detection in rainy images. Final Year Project (FYP), Nanyang Technological University, Singapore. https://hdl.handle.net/10356/149815||Abstract:||With advancements in technology, artificial intelligence (AI) has seeped into our daily lives through innovations such as digital assistants like Amazon’s Alexa, and autonomous vehicles which do not require a human driver. Although autonomous vehicles have seen improvements over the last few years, their lane detection accuracy can still be affected if they experience low visibility. In this project, the deep learning approach will be explored to improve lane detection accuracy in conditions with poor visibility, such as rain. Computer vision techniques, such as masking and regression, would be implemented and evaluated alongside a deep learning architecture to solve this problem. Vanishing point estimation would be used as an auxiliary task to further improve the lane detection process. In this project, detectors operating on different deep learning frameworks would be tested on lane images in rainy and non-rainy conditions. Results from the different deep learning frameworks would be reviewed for their lane detection efficacy using evaluation metrics. An existing lane detection algorithm, VPGNet, would be used as the foundation for this project. It would be trained and tested in its original deep learning framework, Caffe, for its efficacy in lane detection in rainy and non-rainy environments. VPGNet in Caffe was trained using a non-rainy dataset and a dataset that contained both non-rainy and rainy data. After the completion of training, lane detection was evaluated using the F1 score metrics. After VPGNet was evaluated on the Caffe framework, an improved version of the model was proposed for the PyTorch framework. PyTorch was chosen for its intuitive design as well as ease of setting up. The PyTorch model would attempt to recreate the vanishing point task that was not included in VPGNet’s Caffe code. Other loss functions would also be explored under the PyTorch framework to assess its effect on training and testing performance. Results from both Caffe and PyTorch frameworks would be analysed. VPGNet in the Caffe framework showed remarkable results in non-rainy environments as well as in the different environments that arose from rainy weather, such as limited visibility and reflections on the road surface. The vanishing point task in the PyTorch framework would still require improvements to improve its accuracy.||URI:||https://hdl.handle.net/10356/149815||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Student Reports (FYP/IA/PA/PI)|
Updated on May 16, 2022
Updated on May 16, 2022
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.