AI-Supported Rapid Wildlife Detection and Classification at Highway Crossings with Synthetic Thermal Data Creation
Novel two-stage pipeline combining synthetic thermal data generation with real-world classification, achieving 99% detection accuracy and 92%+ species classification for WSDOT's I-90 wildlife corridor.
Problem Statement
Large-scale wildlife monitoring projects, such as the I-90 Snoqualmie Pass East Project, involve extensive networks of motion-activated thermal cameras that generate millions of images and videos annually. Scaling thermal camera coverage becomes impractical without automation, largely due to the lack of effective computer vision models for thermal imagery.
The core challenge is the limited availability of annotated thermal animal data, which prevents training accurate detection and classification models. Traditional monitoring methods are time-consuming and costly, with over 172,000 motion-triggered videos/images backlogged across just two overpasses in a three-month period.
Technical Implementation
Synthetic Data Generation
- • Novel morphing workflow using PlantCV API
- • 26,000 annotated optical images from COCO dataset
- • Domain conversion preserving all annotations
- • Gaussian blur for realistic thermal simulation
- • Largest thermal animal dataset in public domain
Two-Stage Pipeline
- • YOLOv8 detection model on synthetic thermal data
- • YOLOv8 classification model on real thermal data
- • Sequential filtering and species identification
- • Real-time inference optimization
- • Scalable deployment across camera networks
Research Methodology
Data Collection & Synthesis
- • Thermal imagery collection from field deployments
- • Synthetic data generation for model training
- • Multi-species wildlife dataset compilation
- • Environmental condition variation modeling
- • Cross-validation across different camera locations
Model Development
- • Deep learning model architecture optimization
- • Real-time thermal image processing
- • Species classification algorithms
- • False positive reduction techniques
- • Model compression for edge deployment
Results & Performance
The system was tested on 911 thermal video clips from WSDOT motion-activated cameras at the 61.5 Overcrossing South site along I-90. The detection model achieved 99.05% average detection accuracy across deer/elk, bobcat, and coyote, while filtering 98.63% of false positives from motion noise.
Detection Performance
- • Deer/Elk: 100% detection accuracy (340/340)
- • Bobcat: 98.24% detection accuracy (261/270)
- • Coyote: 99.34% detection accuracy (155/157)
- • False Positives: 98.63% filtered (142/144)
- • Average detection accuracy: 99.05%
Classification Performance
- • Deer/Elk: 97.35% classification accuracy
- • Bobcat: 92.72% classification accuracy
- • Coyote: 92.90% classification accuracy
- • Two-stage pipeline optimization
- • Real-time species identification
Technologies Used
Conference Presentations
AI-Supported Rapid Wildlife Detection and Classification at Highway Crossings with Synthetic Thermal Data Creation
Authors: Vedant Srinivas, Joshua Zylstra, Mark Norman, Fraser Shilling
Affiliations: Road Ecology Center, UC Davis; Stanford University; Jacobs; WSDOT
Conference: International Conference on Ecology and Transportation (ICOET) 2023
Computer vision model for automated wildlife monitoring using thermal imagery. The system addresses the challenge of processing millions of images from motion-activated cameras by implementing automated detection and classification. Currently deployed in a WSDOT data center for real-time processing, with plans to expand across additional wildlife crossings.
Automated Wildlife Monitoring on Wildlife Crossings
Authors: Vedant Srinivas (Stanford University), Mark Norman, Josh Zylstra, Fraser Shilling
Conference: Wildlife Society Western Section (WSTWS) 2023 - Student Paper
Computer vision model achieving 100% precision and 98.80% recall on 270 videos of 813 deer from wildlife crossings on the I-90 corridor. The model is currently deployed in a WSDOT data center for real-time classification of thermal camera footage.
Future Work
The system is currently being deployed on cameras at two overcrossings, with plans to expand to 12 additional sites. Future work will focus on incorporating additional wildlife species common to the Pacific Northwest, implementing feedback loops from deployed systems, and optimizing for embedded edge computing.
Beyond detection and classification, future pipelines could extract behavioral metrics, integrate temporal trends, and support deeper ecological analysis. The framework offers a path toward more automated, accurate, and adaptive wildlife monitoring systems across broader ecological applications.