news

       Content generated by artificial intelligence may not fully reflect the essence of the information. Please verify important information. Learn more.
       Editor’s note: This is the final article in our NVIDIA DRIVE Labs series, which explores the engineering challenges of autonomous vehicles and how NVIDIA DRIVE is addressing them. Click here to read all of our automotive-focused articles.
       Detecting lane markings and road edges is crucial for the development of autonomous vehicles—lane marking recognition is the foundation of systems like lane departure warning, which helps human drivers prevent lane departures. In addition to detecting lane markings, autonomous vehicles must also recognize other road signs (such as arrows or the word “stop”) and vertical landmarks, which helps accurately position the vehicle on a high-resolution map.
       This week at DRIVE Labs, we’ll be presenting an evolution of the LaneNet DNN neural network, which already boasts high accuracy and stability in road marking detection. This evolution upgrades it to the highly accurate MapNet DNN neural network. This enhancement not only expands the detection categories to include road markings and vertical landmarks (such as utility poles), but also increases the detection range. Furthermore, end-to-end detection improves processing efficiency, enabling faster data processing directly in the vehicle.
       The MapNet DNN model included in NVIDIA DRIVE software version 10.0 is capable of detecting painted lane markings (solid/dashed lines, intersection entry/exit lines, road edges), painted road markings (e.g., arrows, stop signs, and high-occupancy vehicle lane markings), and vertical poles (e.g., road signs and lampposts).
       To achieve high-precision road markings and vertical landmark detection, MapNet DNN uses a basic real-world data encoding technique similar to its predecessor, the high-precision LaneNet. This encoding method prevents the loss of high-resolution visual information during convolutional DNN processing and is independent of orientation and location. In addition to generating sufficient redundancy to preserve rich lane information, it can also be easily extended to preserve information about road markings and arbitrary landmarks (e.g., utility poles, which in this case can be considered “perpendicular lane lines”).
       We also noticed that MapNet’s high-precision system can accurately detect road markings even when some elements are missing. When a single lane contains both solid and dashed markings, MapNet treats the markings as solid, ensuring safe driving.
       MapNet can also detect road edges, which is particularly useful in situations where clear lane markings are not available, and continuously track the transition from solid to dashed lane lines.
       We’ve found that even in visually challenging conditions, such as cracks in the road, stains on the asphalt, and strong shadows from trees or vertical landmarks, MapNet consistently recognizes lane markings and road edges. Furthermore, MapNet can recognize road markings in multiple languages.
       The latest deep neural network model, MapNet, currently in development, is trained for end-to-end detection of road markings and landmarks, significantly reducing the complexity of post-processing the DNN’s raw results to produce continuous geometric output. Rapid acquisition of information directly in the vehicle is crucial, as it provides low-latency perceptual input for longitudinal and lateral planning and control functions.
       Furthermore, the highly accurate road marking and landmark detection data provided by MapNet can serve as input for autonomous vehicle mapping and positioning functions. The ability to detect vertical landmarks (such as power line pylons) is also particularly important for achieving accurate longitude positioning results.