Implementing Machine Learning Based Image Recognition for Animal Detection

Today, ecologists are aiming to improve the monitoring and safekeeping of animal wildlife. In an effort to make data more reliable for wildlife researchers, ecologists are increasingly turning to environmental measurement and data analysis. Computing systems can help automate this process to provide more reliable data in a completely automatic manner. This project focuses on the use of computer vision combined with machine learning to detect animals in photos taken by motion-triggered camera traps. With the use of open source, pre-trained Convolutional Neural Networks a machine learning model can “learn” to recognize wild animals to some extent. After setting up the CNN framework trained on the default ImageNet dataset, we ran the classifier on 12210 images and found the model to be inaccurate. We further extended the accuracy and range of the recognition by re-training this model with labeled images from the motion-triggered camera traps. The re-training phase requires a set of labeled animal images that the model uses to create its recognition database. The next step is to use the recognition data and combine it with the image's metadata – that is, extra information such as time, date, and temperature. Combining the recognition data with its metadata can help us learn more about how animals are affected by the environment. Further data analysis can help predict endangered animal species by noticing unusual trends. The information extracted from the images can help us address important scientific concerns, while providing a unique view into the hidden world of wildlife. 

Project Mentor: Nevena Golubovic

Faculty Advisors: Chandra Krintz, Rich Wolski