SEMANTIC SEGMENTATION IN SELF-DRIVING CARS USING U-NET ARCHITECTURE ON CITYSCAPE DATASET

Semantic segmentation has been one of the most researched topics in the field of computer vision in recent years. This study was conducted using U-Net architecture in the context of self-driving cars on a cityscape dataset. The dataset is an urban scene image that contains all scene scenarios in a typical city. It includes 5,000 high-quality finely annotated pixel-level images gathered from 50 cities over various seasons. The U-Net model uses a pre-trained RestNet101 for its encoder for feature extraction and has skip connections between the encoder and the decoder with RELU activation. The skip connection helps to retain spatial information after down sampling, this enables the model to combine deep layers and fine-grained features. The model achieved 88\% accuracy, 80\% pixel accuracy, 85\% precision, 84\% recall and 84.49\% F1 score metric. The model was trained for 75 epochs of 2 hours and 30 minutes of training time on the cityscape dataset. The model has shown good performance by achieving high accuracy and addressing class imbalance with augmentation techniques and weighted categorical loss in the context of autonomous driving. Theproposed U-Net model with RestNet101 encoders achieved high accuracy compared to VGG16 and ResNet50 in a typical complex scene environment in the cityscape dataset.