Pavement Condition Index Model

Model by Booz Allen

This model calculates the Pavement Condition Index (PCI) of a section of road based on an input image. This model accepts JPEG images of any size, though minimum recommended resolution is a width of 1500px for a 2 lane road and taken from a top down aerial perspective. From this image, an ensemble of models detects and isolates the damage in the image, and from them calculates a final PCI score. This is output in a JSON format.

  • Description

    Product Description

    PERFORMANCE METRICS:

    58% F1 Score

    F1 Score is the harmonic mean of the precision and recall, with best value of 1. It measures the balance between the two metrics. Further information here.

    This model was trained on an internal dataset of aerial drone imagery collected in and around Los Angeles, California. This model is adept at calculating a PCI that is in line with expert analysis, however performance degrades with low resolution imagery. This model performs best without obstructions such as cars on the road, and with a minimum resolution of 1500px width for a two-lane road. Lanes with a larger number of lanes should be scaled accordingly as it’s vital the resolution is high enough to capture cracks in the pavement. For fast inference time, this model works best on a GPU.

    OVERVIEW:

    This model uses the FCN architecture to perform several different functions in order to arrive at a final PCI. The first model in the pipeline locates and segments out the asphalt from the rest of the image in order to ensure we’re not detecting cracks on sidewalks or other infrastructure that isn’t pertinent. Then, a second model detects all the cracks and spalling in the extracted asphalt. The classification, orientation, width, and length of each of these cracks is determined, and then factored into the final PCI value

    TRAINING:

    This model was trained on aerial drone imagery on a collection of roads within Los Angeles. Approximately 500 labeled images were hand labeled, verified, and used to train the model. The model was trained using stochastic gradient descent with a learning rate of 0.01 and momentum of 0.9 for 10 epochs.

    VALIDATION:

    Validation was performed on a 20% holdout portion of the original data that was collected.

    INPUT SPECIFICATION

    The input(s) to this model must adhere to the following specifications:

    Filename Maximum Size Accepted Format(s)
    input.jpg 10M .jpg

    OUTPUT DETAILS

    This model will output the following:

    Filename Maximum Size Format
    results.json 1M .json