Pothole Finder Model

Autoencoder-based system to take accelerometer, gyroscope, location, and speed data from an Android device and find out where the phone experienced a pothole.

File descriptions and explanation

standardiser.py

Standardises the readings obtained from the device to every third of a second. Required so that devices with lower sensor polling rates do not give a differently sized input to the model.

Usage

$ python standardiser.py --input <File path to raw readings file> --output <File path to desired output file>

If you wish to change the standardising rate from every third of a second to something else, change the following line as indicated:

# calculating averages for every third of a second
i = 1
old_j = 0
j_limit = 0
indexes.insert(0, 0)
for index in indexes[1:]:
    old_j = indexes[i - 1]
    for n in range(3): # change 3 to n here for readings to be standardised to every nth of a second`

pothole_predictor.py

Trains autoencoder to recognise patterns in the readings of a phone being used over a road with no potholes.

Usage

$ python pothole_predictor.py --input <File path to standardised training dataset> --mode <1 for considering only accelerometer readings, 2 for considering only gyroscope readings, 3 for both>

Explanation

The model takes in all input features, compresses the same to a single-dimensional layer that represents most essential information in the training dataset. This is then expanded back to the dimensionality of the entire input dataset. Using the mean-squared-error as a loss function, thresholds are established after the minimisation of MSE. These thresholds are then used for the tester.py script, which is explained below.

tester.py

Uses model .h5 file generated by pothole_predictor.py to output results as per thresholds established according to analysis of the plot output during execution of pothole_predictor.py.

Usage

$ python tester.py --input <File path to standardised test dataset> --mode <1 for accelerometer testing, 2 for gyroscope testing, 3 for both>

Explanation

This line deserves most attention:

if 6.49795 <= error_df.iloc[i][0] <= 13.0633:

The bounds on the error values isolate the pothole locations to what I term as moderate outliers. The data input into the model can be split into 3 categories according to the reconstruction error values:

1. Least values (0, 6.49795): Not potholes
2. Moderate outliers (6.49795, 13.0633): Potholes
3. Erroneous outliers (> 13.0633): User behaviour-influenced changes in readings (example: picking up a call while logging a trip)

model_freezer.py

Converts the .h5 files generated by Keras to .pb file intended for use in the Android app Road Quality Audit.

Must be executed before deploying files to the app (TensorFlow for Android cannot read .h5 files)

Usage

$ python model_freezer.py