NVIDIA Launches LaneNet DNN Model for High-Precision Detection of Road Signs and Landmarks

2019-12-07

Lane detection and road edge detection are crucial for the development of autonomous vehicles. Lane detection provides power to systems such as lane departure warning, helping drivers avoid lane deviation. In addition to detecting lane line information, autonomous vehicles also need to detect other road markings (such as arrows or STOP signs) and vertical landmarks that help accurately position the vehicle.


In Nvidia's driving lab project, they introduced the LaneNet DNN, a lane detection neural network model that can accurately and stably detect painted lane lines on the road and input them into the MapNet DNN, a map neural network model. This development includes the addition of detection categories, including not only lane detection but also road markings and vertical landmarks. It also improves processing efficiency through end-to-end detection, providing faster in-vehicle inference capabilities.


Painted road markings (such as arrows, STOP signs, and other frequently occurring lane markings) as well as vertical landmarks (such as road signs and streetlights).In order to perform accurate detection of road markings and vertical landmarks, MapNet DNN utilizes the advanced ground truth data encoding technology of its predecessor, the high-precision LaneNet. This encoding prevents the loss of high-resolution visual information during convolutional DNN processing. In addition to creating enough space to retain rich lane line information, it can also easily expand to retain information about road markings and landmarks of arbitrary shapes.


They also observed that even in cases where some painted markings are missing, the high-precision MapNet can still provide accurate detection of road marking shapes. If solid and dashed lane markings are placed on the same lane, MapNet intentionally treats the lane markings as solid lines to support safe driving.


MapNet can also detect road edges, which is particularly useful when there are no clear painted lane markings, and it can always detect the transition from solid to dashed lane markings. It can stably detect lane lines and road edges even in the presence of visual obstacles, including road cracks, asphalt stains, and rough shadows projected by trees or vertical landmarks. In addition, MapNet can detect road text markings in different languages.


The latest version of the MapNet DNN model currently under development has been trained to perform end-to-end detection of road markings and landmarks, greatly reducing the complexity of post-processing the original DNN results into continuous geometric outputs. Fast in-vehicle inference is crucial as it provides low-latency perception input for longitudinal and lateral planning and control functions.


Furthermore, the high-precision road marking and landmark detection results provided by MapNet can be used as input for the mapping and localization functions of autonomous vehicles. The ability to detect vertical landmarks (such as magnetic poles) is particularly advantageous for obtaining accurate longitudinal positioning results.

RECENT AI NEWS

RECENT AI TOOLS