Deep Learning Based Multi-Sensor Fusion and Uncertainty Estimation with Continual Learning for Dynamic End-to-end Lane Detection

Autonomous vehicle technology demands advanced lane detection (LD) techniques adaptable to complex road scenarios. Traditional LD systems relying on static inputs often fail in real time. The proposed research introduces a deep learning model integrated with a sensor fusion strategy in an IoT- based vehicular perception framework to address these limitations. The proposed end-to-end uncertainty estimation with continual learning network (E2E-UCNet) utilizes data collected through cameras from diverse real-time environments in Andhra Pradesh and Tamil Nadu, India. In the absence of physical range sensors, depth cues similar to LiDAR and Radar data are generated from monocular RGB images using MiDaS monocular depth estimation and incorporated into the E2E-UCNet model to improve feature representation for dynamic lane prediction. The Multi-Sensor Fusion with Uncertainty Estimation (MSF-UE) enhances E2E-DLN reliability through Bayesian deep learning. Further enhanced by the continual learning semantic segmentation network (CL-SSN), the framework achieves 99% pixel-wise accuracy, 0.51 IoU, and a Dice coefficient of 0.67 under varying road and weather conditions. The dynamic continual learning for long-term adaption (DyCLA) framework progressively updates the model with new data, ensuring consistent lane detection under changing traffic conditions. These findings aim to improve self-driving safety and reliability for autonomous vehicle applications.