About This Module

The Parking Space Perception module detects available parking spaces from images taken by a single camera. It also identifies entry lines into parking spaces. The NVIDIA proprietary DNN named ParkNet returns an array of objects each constituting an available parking space. A parking space object is represented by the coordinates of four corners of a quadrilateral shape. One side of such quadrilateral shape is classified as an entry line. The ParkNet DNN infers parking spaces in the image coordinates. Using the extrinsic parameters of the camera within the sensor-rig and a flat-ground assumption, this module converts the image-based coordinates of the quadrilateral shape into 3D-world space coordinates. This can then be used by other modules within the DriveWorks software stack.

The included data structures are as follows:

dwParkObject, which defines a single parking space object.
dwParkDetection, which defines the output of the inference from a single camera frame.
dwParkNetDetectorParams, which defines parameters of the parking space detector module such as camera extrinsics, image resolution and others.

Additional Information

Warning

Currently, the parking space detector module has several limitations that could affect its use. These limitations are:

The module processes a single frame at a time. Processing images from several overlapping cameras is completely independent - not merging of detections is performed.
The current module does not batch many images for performance speed-up, which may be an inconvenience for a multi-camera system.
The module assumes a flat-ground world. If the current scenery has uneven ground, the 3D prediction will exhibit a 3D-location error even with perfect detection in the image space.
See additional warnings affecting detector performance which relate to the ParkNet-DNN training: ParkNet.

Relevant Tutorials

APIs

Sample App

Parking Space Detection Sample