Bookmark and Share

Computer Vision Metrics: Chapter Eight (Part C)

Register or sign in to access the Embedded Vision Academy's free technical training content.

The training materials provided by the Embedded Vision Academy are offered free of charge to everyone. All we ask in return is that you register, and tell us a little about yourself so that we can understand a bit about our audience. As detailed in our Privacy Policy, we will not share your registration information, nor contact you, except with your consent.

Registration is free and takes less than one minute. Click here to register, and get full access to the Embedded Vision Academy's unique technical training content.

If you've already registered, click here to sign in.

See a sample of this page's content below:

For Part B of Chapter Eight, please click here.

Bibliography references are set off with brackets, i.e. "[XXX]". For the corresponding bibliography entries, please click here.

Calibration, Set-up, and Ground Truth Data

Several key assumptions are made regarding scene set-up, camera calibration, and other corrections; we summarize them here:

  • Roadway depth surface: Depth camera is calibrated to the road surface as a reference to segment autos above the road surface; a baseline depth map with only the road is calibrated as a reference and used for real-time segmentation.
  • Device models: Models for each car are created from manufacturer’s information, with accurate body shape geometry and color for each make and model. Cars with custom paint confuse this approach; however, the shape descriptor and the car region depth segmentation provide a failsafe option that may be enough to give a good match—only testing will tell for sure.
  • Illumination models: Models are created for various conditions, such as morning light, daylight, and evening light, for sunny and cloudy days; illumination models are selected based on time of day and year and weather conditions for best matching.
  • Geometric model for correction: Models of the entire FOV for both the RGB and depth camera are devised, to be applied at each new frame to rectify the image.

Pipeline Stages and Operations

Assuming the system is fully calibrated in advance, the basic real-time processing flow for the complete pipeline is shown in Figure 8-5, divided into three primary stages of operations. Note that the complete pipeline includes an image pre-processing stage to align the image in the FOV and segment features, a feature description stage to compute shape and color descriptors, and a correspondence stage for feature matching to develop the final automobile label composed of a weighted combination of shape and color features. We assume that a separate database table for each feature in some standard database is fine.

No attempt is made to create an optimized classifier or matching stage here; instead, we assume, without proving or testing, that a brute-force search using a standard database through a few thousand makes and models of automobile objects works fine for the ALPHA version.

Note in Figure 8-5 (bottom right) that each auto is tracked from frame to frame, we do not define the tracking method here.