Enhancing Level 5 Autonomous Driving: With Multi-Model Integration and Real Time Prioritization

Recently, many self-driving cars and technologies have been developed. However, level 5 driving has not yet been achieved. The transition from Level 4 to Level 5 demands significant advancements due to the increased complexity and autonomy required. To address this “last-mile problem” of self-driving technology, this study identifies key challenges in current systems, focusing on accurate and timely steering predictions under conditions such as high speed, low light or nighttime environments, and unclear or blurred road lanes-scenarios where existing models often underperform. This research begins with a baseline steering model and incrementally refines it into a multi-model framework incorporating real-time functionality. Specifically, it integrates a lane prediction model to enhance lane visibility in challenging environments, employs TensorRT for model optimization to reduce processing latency, and utilizes multi-threading and real-time (RT) priority assignments by using preempt-rt to maximize CPU utilization for critical tasks. Overall, this paper analyze the improvements and limitations of each step of steering models and more importantly suggests that we can get closer to level-5 driving through real-time function support.