Citation

Al Mamun, Abdullah (2021) Lane Markings Detection Using Encode-Decode Instant Segmentation Network Algorithm. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

In recent times, many innocent people have suffered from disabilities and sudden death due to unwanted accidents on the roads, which also rivets lots of financial assets. Hence, researchers are incorporating many modern and significant features of ADAS. Lane marking detection is one of the most preliminary and significant ADAS features, which allows the vehicle to maintain the perspective road lane itself. The present researches using Deep Learning techniques have some research limitations from different perspective challenges. The researchers are most commonly facing difficulties in detecting the lane marks due to the environmental effects such as the variant of lights, obstacle, shadow, and curve lanes. For that reason, it also experienced less numerical performance results like accuracy, F1-measure, precision, recall, etc., on the Lane Marking Detection (LMD). Computational speed is also another concern in LMD techniques. Therefore, this thesis proposed a DL technique named Encode-Decode Instant Segmentation Network (EDIS-Net) to detect the lane markings under different environmental conditions with high-performance results. The framework is based on the E-Net architecture incorporating a combination of discriminative and cross-entropy losses. The encoded section was split into two stages: binary and instant segmentation to extract the information about the lane pixels and pixel position. DBSCAN is used to interface the predicted lane pixels to have the final output. The framework was trained on the Tusimple dataset with data augmentation technique and tested on three datasets such as Tusimple, CalTech, and local datasets. The model has achieved 97.39% accuracy, 68.2% F1 score, 98.01% precision, 97.29% recall, 3.421% False Positive Score, and 1.359% False Negative Score on the Tusimple dataset. Again, it has obtained 97.07% and 96.23% average accuracy on CalTech and the local dataset, respectively. All the experimental results are also compared with existing LMD techniques, where the EDISNet showed promising results. It is expected that this research will bring a significant contribution to the lane marking detection field and serve society by resisting sudden accidents.

Item Type:	Thesis (Masters)
Additional Information:	Call No: Q325.73 .A23 2021
Uncontrolled Keywords:	Deep learning (Machine learning)
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	24 Feb 2023 06:30
Last Modified:	24 Feb 2023 06:30
URII:	http://shdl.mmu.edu.my/id/eprint/11150

Downloads

Edit (login required)