ttenneb / duplicate_detector

Duplicate Detector Model

This project contains two scripts: generate.py and train.py. The first script generates the feature used by the model from trespassing data. The second script trains a random forest to predict if a given trespassing incident is a duplicate.

Data and Feature Generation - generate.py

Required Packages:

• csv
• opencv
• numpy
• scikit-image
• scikit-learn
• ast

The feature generation script takes the name of the dataset as a parameter. The dataset must be in csv format. A csv file with the Thomasville North trespassing data is included. After installing the required packages and adding a csv file to this directory the script can be run with the command below:

python generate.py "Thomasville N.csv"

This script will most likely take a couple of hours to generate the features.

The following features will be generated in a new csv file:

Trajectory difference:

There are 4 features that model the difference between two events trajectories. First, the summed difference of trajectory points between the 2 events. Second, The log of the summed difference. Third, the difference between the events first trajectory points. Fourth, the log of the start difference.

Time:

The seconds elapsed between the two events is the only feature in this category.

Linear Trajectory:

The trajectories of both events are modeled using linear regression. The coefficients and intercepts are used as features.

Previous Duplicates:

Since duplicates are likely to happen consecutively the model takes to features to model past duplicates. First, a boolean value representing whether the last event was a duplicate. Second, an integer value representing the current number of consecutive duplicates.

SSIM score:

The Structural Similarity index between the last frame of the first event's video and the first frame of the suspected duplicate's video.

Model Training - train.py

Required Packages:

• csv
• opencv
• numpy
• pandas
• scikit-image
• scikit-learn
• matplotlib

A csv file name "data.csv" must be included in the directory.

This script only trains a model for testing purposes.

Training the model is very straight forward. Simply run the command below:

python train.py

An ROC curve and accuracy metrics will be displayed once the model has been trained.

Selecting a Model

Some models perform much better than others. Accuracy is not helpful for selecting a model in this situation since false positives are much worse than false negatives. Therefore, the ROC curve needs to be used to select a model. Two example ROC curves are shown below. Both curves are from models with the same architecture and data, but example 2 is a much more useful model. Example 2 is more useful because it has a threshold that has almost no false positives while still catching almost 50% of the duplicates.

Example 1:

Example 2:

Planned

• Optimize hyperparameters.
• Finish checking Thomasville North dataset for unlabeled duplicates (many still exist).
• Test new model architectures.
• Test model with uncertain previous duplicate features.
• Add additional SSIM features.
• Add trajectory direction feature.