Deep Learning-Based Lane-Keeping Assist System for Self-Driving Cars Using Transfer Learning and Fine Tuning

This paper presents an advanced lane-keeping assistance system specifically designed for self-driving cars. The proposed model combines the powerful Xception network with transfer learning and fine-tuning techniques to accurately predict the steering angle. By analyzing camera-captured images, the model effectively learns from human driving knowledge and provides precise estimations of the steering angle necessary for safe lane-keeping. The transfer learning technique allows the model to leverage the extensive knowledge acquired from the ImageNet dataset, while the fine-tuning technique is utilized to tailor the pre-trained model to the specific task of steering angle prediction based on input images, enabling optimal performance. Fine-tuning was initiated by initially freezing the pre-trained model and training only the Fully Connected (FC) layer for the first 10 epochs. Subsequently, the entire model, encompassing both the backbone and the FC layer, was unfrozen for further training. To evaluate the system's effectiveness, a comprehensive comparative analysis is conducted against popular existing models, including Nvidia, MobilenetV2, VGG19, and InceptionV3. The evaluation includes an assessment of the operational accuracy based on the loss function, specifically utilizing the Mean Squared Error (MSE) equation. The proposed model achieves the lowest loss function values for both training and validation, demonstrating its superior predictive performance. Additionally, the model's performance is further evaluated through extensive real-world testing on pre-designed trajectories and maps, resulting in the minimal deviation of the steering angle from the desired trajectory over time. This practical evaluation provides valuable insights into the mode's reliability and its potential to effectively assist in lane-keeping tasks.