Trail Camera Classification

Inspiration

Trail cameras take a picture anytime they detect motion. This means they get a lot of false positives where there is either no animal in the picture, or it is an animal that they are not tracking. My goal is to save DNR officers and Hunters time by classifying there images and allowing them to search by the classifications.

What it does

Currently my model will classify deer versus non deer images. When I gather a larger dataset I will be able to classify coyotes, bears, racoons, and turkeys. My product will allow them to search for individual animals in there image sets and even filter by time the pictures were take.

How I built it

I built my project by retraining the VGG16 ImageNet classifier on my dataset using numpy and keras on top of tensorflow in python.

Challenges I ran into

I ran into challenges with other classifier architectures. What I found was my accuracy plateauing at 15%. The mentors at Caterpillar helped me work through a lot of issues. Getting stuck at low accuracy allowed me to try different image pre-processing and models to trouble shoot my issues. I learned a lot about machine learning through this.

Accomplishments that I'm proud of

I am proud that I stuck with it and now have accuracy in the high 80's.

What I learned

I learned new methods of image pre-processing as well as how to import weights of pre-trained models. I also learned how to modify the architecture of pre-trained models and retrain them to get higher accuracy on my own data.

What's next for Trail Camera Classification

I plan to deploy my model to a desktop app to allow users to quickly and effectively classify their own deer images. I also will be collecting more trail camera images to increase the size of my data set. This will allow me to classify other wildlife and provide more use to DNR officers.

Built With

• keras
• numpy
• pandas
• python
• tensorflow