Computer Vision Automation of Wildlife Monitoring on WSDOT’s I-90 Snoqualmie Pass East Project

ICOET 2025
Authors
Vedant Srinivas, Stanford University
Dr. Fraser Shilling, The Road Ecology Center at UC Davis
Mark Norman, Washington State Department of Transportation
Josh Zylstra, Jacobs Engineering
Abstract

Wildlife monitoring is crucial for informing conservation efforts, providing valuable insights into animal infrastructure use, movement patterns, and population dynamics. Traditional monitoring methods that include motion activated cameras and manual analysis are not only time-consuming but also expensive and prone to human error (e.g., data transcription). 

The technology addressed in this project was focused on monitoring wildlife crossings, specifically those being built as part of the I-90 Snoqualmie Pass East Project.  The wildlife monitoring program for this project involves 15 networked, motion-activated thermal cameras, generating over two million images and videos annually. Monitoring and documenting this data currently requires a dedicated full-time biologist. False positive detections are becoming more of a challenge as vegetation within and around the existing structures matures. Additionally, as the project progresses and new crossing structures are added, the camera count is anticipated to double, rendering comprehensive and timely coverage impractical and costly without automation. The lack of automation has been attributed to the absence of effective computer vision models for thermal imagery.

The solution was the creation of a computer vision model capable of classifying thermal animal footage with high accuracy. The model was trained on simulated thermal data, created through a custom image morphing algorithm, as well as real thermal data. The image morphing algorithm was developed to convert large databases of color or black and white optical camera images into simulated thermal images that can be used for thermal model training. The purpose of the model was to remove false positive detections from the motion detecting camera, and classify footage by species (deer, elk, otters, pumas, etc.). This classified footage was documented to a metadata file including details about the location, date/time, and classification, including species type and number of animals. 

The model was trained on a dataset of 26,000 simulated thermal images across 10 animal classes as well as 720 images of deer collected from WSDOT cameras. The model will continue to be refined with data collected from the field and annotated by volunteers. Using testing data not used in model training, the newly trained model achieved a precision of 100% and a recall of 98.80% on real-world deer crossing data (270 videos of 813 deer) from the Keechelus Lake Wildlife Overcrossing on the I-90 corridor. A precision of 100% means there were no false positives, which solves the key problem with the current motion activated cameras used by WSDOT. The model has been deployed on a workstation in a WSDOT data center for real-time classification of footage from cameras on one overcrossing. WSDOT’s plan is to scale to 12 more cameras on crossings as they are built and monitored.