dkrautha / cats-ece

C.A.T.S - Camera Assisted Tracking System

Stray Cat Tracking in the AI Age

Abstract

The growing population of stray cats presents a significant challenge for communities across the United States, with an estimated 1.3 million roaming the streets of New Jersey alone, and a whopping 70 million nationwide. Addressing this issue has typically relied on labor-intensive and costly efforts, with Trap, Neuter, Release (TNR) programs emerging as the most progressive and humane approach. However, the success of TNR is reduced by the nomadic nature of stray cats, requiring workers to manually locate and trap large populations. In a revolutionary stride towards more efficient and cost-effective solutions, the emergence of the Camera Assisted Tracking System (C.A.T.S.) promises to transform the landscape of stray cat management. C.A.T.S. is an innovative AI-based camera system designed to effectively identify and track stray cat populations within a given community. This new technology aims to automate the tracking process, greatly reducing costs and simplifying the difficult work undertaken by animal control groups. At the core of C.A.T.S. is its AI algorithms, capable of recognizing and analyzing stray cat patterns with great accuracy. The system utilizes well placed cameras equipped with advanced image recognition capabilities, ensuring thorough coverage of target areas. As stray cats traverse their territories, C.A.T.S. captures their movements, creating a detailed and real-time database of their locations.

Our Contributions

This is a continuation of the a previous senior design project, which you can find here. Credit for the original idea, fantastic name, and any commits not made by David Krauthamer, Brandon Boutin, and Robert Plastina, go to the original authors.

Our goal as a team was to improve on the previously created model by optimizing both it's speed and accuracy. We wanted to use Google Coral hardware to accelerate the model to obtain the speed increase, hence why we switched to using TensorFlow (Lite). To increase the accuracy we trained the model to identify not just cats, but other animals which could be easily mistaken by a computer for a cat. The expectation was that this would reduce the false positive rate, and increase the accuracy of the model overall.

The poster we created for the Stevens Innovation Expo can be found here.

The website we created for Senior Design (D7 and D8) can be found here.

Dataset Sources

• Cat and Dog breeds: https://www.kaggle.com/datasets/tanlikesmath/the-oxfordiiit-pet-dataset
• Raccoons: https://universe.roboflow.com/raccoondetector/raccoon-finder
• Squirrels: https://universe.roboflow.com/project1373/squirrels-faces
• Foxes: https://universe.roboflow.com/fox-ewljn/fox-oljyb
• Coyotes: https://universe.roboflow.com/david-casas/coyote-uuhaf

Each of the non-cat and dog datasets is available in tfrecord format, and that's what we used for training. The cats and dogs tfrecord needed to be generated, we the modified instructions from here in order to generate it.

From what we've been able to find online, merging multiple tfrecords together into a single large tfrecord isn't the best practice, however it's what we chose to do for simplicity. The script used to combine the tfrecords is combine_tfrecord.py. The reason for combining the tfrecords into one is it allowed us to use the a script provided from Google themselves here.