Paper #4

Optimization of Road Detection using Semantic Segmentation and Deep Learning in Self-Driving Cars

Mohammed Sameeh Hammoud and Sergey Lupin

Abstract: Robust and accurate road detection is an essential part of Automatic Driver Assistance Systems (ADAS). Self-driving Cars have the capability to revolutionize the way we travel, making transportation safer, more effective and more available to all. With the ability to navigate roads without human intervention, self-driving cars can reduce the number of accidents caused by human error, eliminate the demand for drivers to be behind the wheel and make it easier for people who can’t drive, such as the elderly or disabled people to get around. In addition, self-driving cars can enhance traffic flow by reducing congestion and optimizing routes, eventually saving time and reducing emissions. As technology continues to advance, self-driving cars are poised to transform the transportation industry and change the way we think about mobility. In this work, a convolutional neural network-based deep learning to achieve road detection based on image segmentation to be applied in self-driving cars. In addition to our proposed network, multiple experiments were conducted to investigate the impact of different deep-earning architectures on performance. A public dataset called the KITTI road dataset is used to train and validate the model. The images were down-sampled from 1224x370 to 256x256. We compared our model’s performance with the performance of popular deep learning architectures such as Unet and LinkNet A transfer learning technique is used while training the models based on network weights trained on the famous dataset ImageNet, including popular architectures such as ResNet, VGG, SeresNet and EfficientNet. The results show that our model achieves an F1-score of 0.9909, outperforming Unet and LinkNet architectures. In the second place, the best results were obtained based on Unet and ResNet50 with an F1-score of 0.9904.

Keywords: Advanced Driver Assistance Systems; Computer vision; Deep Learning; Image segmentation; Mobile Robots; Navigation; Road Detection; Self-Driving Cars.