DriveWorks SDK Reference 3.0.4260 Release For Test and Development only
PathNet

This module provides the APIs to initialize, query, and run the NVIDIA proprietary pathnet detection neural network: PathNet. The included data structures include the following. dwPathPositionType, which lists the types of paths the network will detect. dwPathNetModel allows users to choose from a default or user trained PathNet model. And dwPathNetParams, which defines the PathNet model variant with specific precision and processor optimization to be loaded.

Inputs

PathNet consumes FP16 planar frames with a resolution of 960x604 pixels from AR0231 cameras (revision >= 4) and processes only the bottom 960x400 pixels (cropping is handled in the code). PathNet has been trained specifically on the front camera location with a 60° field of view.

Outputs

PathNet detects between zero and three paths as polygons with a left and a right edge.
The following classes are currently supported:

• Ego path (path of the ego vehicle)
• Left adjacent path (path adjacent to the ego path on the left)
• Right adjacent path (path adjacent to the ego path on the right)

Pathnet also produces a confidence that a given path contains traffic traveling in the opposite direction as the ego vehicle. This is only valid for the left and right adjacent paths as, by definition, traffic in the ego path is headed in the direction of the ego vehicle.

Detections coming from PathNet are associated with:

• a confidence value in the [0,1] range
• a confidence of opposite traffic direction value in the [0,1] range

Warning
Currently, the PathNet DNN has limitations that could affect its performance:
• A single common threshold is used for all paths to determine whether a path should be returned from the detector. This can result in paths flickering if the threshold is set too high. The recommended level is 0.5, but this can be adjusted with the confidenceThreshold field in the configuration file.
• PathNet is trained primarily on roads where drivers drive on the right-hand side of the road. As a result, the accuracy of the traffic direction confidence in the right adjacent path may be reduced.
• PathNet has diminished performance on curved roads that have a significant elevation gain, such as cloverleaves, due to imperfections in the projection to three-dimensional space.