Real-Time Autonomous Driving in High-Illumination Conditions Using Kalman Filter

Advanced Driver Assistance Systems (ADAS) are becoming more and more important as autonomous driving technology advances quickly. To guarantee safe driving in a variety of environmental scenarios, including high-illumination conditions, precise lane detection is especially crucial. This study proposes a reliable lane detection method for high-illumination conditions. The proposed method builds a diverse lighting dataset and applies Kalman filter-based post-processing to correct missed or inaccurate detections. Furthermore, the Xilinx Deep Learning Processing Unit (DPU) was used to optimize and implement the YOLOv3-tiny model for faster processing to enhance real-time inference performance. The proposed method was validated through both GAZEBO simulations and real-world driving experiments. Under 80% brightness conditions, an average of 0.9 lane deviations and 5.1 lane intrusions were observed per 10 laps. Real-time processing requirements were satisfactorily met by the DPU-based implementation, which averaged 82.89 FPS during inference and 52.57 FPS throughout complete system operation.