Abstract
Wildlife-vehicle collisions (WVCs) pose a significant threat to public safety and biodiversity. This paper presents the Smart Wildlife Warning Sign (SWWS), an IoT-integrated solution that leverages sensor fusion and edge-based deep learning to mitigate these risks. By combining infrared and microwave detection with on-device image classification using a VGG16 CNN, the system provides real-time driver warnings and serves as a distributed data collection node for ecological research. Experimental results demonstrate a 95% prediction accuracy when utilizing Laplacian sharpening for image preprocessing.
Introduction
Millions of animals are killed annually in the United States due to wildlife-vehicle collisions (WVCs). Beyond the ecological impact, these incidents result in substantial human injury and property damage. Traditional mitigation strategies, such as static warning signs, often suffer from "sign fatigue," where drivers eventually ignore them. High-cost infrastructure like wildlife overpasses or extensive fencing is often financially unfeasible for widespread deployment.
Furthermore, ecologists lack high-fidelity, real-time data on animal movement patterns, often relying on manual sightings or expensive radio collars. There is a critical need for a dynamic system that can both warn drivers in real-time and gather data on road-crossing events to inform future infrastructure planning.
System Design and Methods
The SWWS is a modular hardware and software platform designed for remote field deployment. The system architecture integrates several key components:
Sensor Fusion
To minimize false positives from environmental noise, the system employs a dual-sensor approach. A passive infrared (PIR) sensor array detects heat signatures, while microwave radar provides depth detection. The system only triggers the camera and warning lights when both sensors confirm the presence of a moving object within a specific range.
Edge AI and Hardware
The core of the system is a Raspberry Pi equipped with a No-InfraRed (NoIR) camera. Upon sensor trigger, the camera captures an image which is processed locally using a VGG16 Convolutional Neural Network (CNN). This edge-computing approach ensures low latency and reduces the need for constant high-bandwidth data transmission.
Communication and Mapping
Data regarding the detected species, timestamp, and location are transmitted via a wireless Ad-Hoc network to a central server. This data is then visualized using the Folium library in Python, creating interactive "bubble choropleth" maps that highlight wildlife activity hotspots and seasonal migration patterns.
Data Preprocessing and Model Training
The VGG16 model was trained on a dataset of over 5,000 images from the US Department of Fish and Wildlife Image Library, focusing on six high-risk species: Black-Tailed Deer, Mule Deer, Rocky Mountain Elk, Roosevelt Elk, Coyote, and Cougar.
A critical finding of this research was the impact of image preprocessing on classification accuracy. Given that wildlife images are often captured in low-light or low-contrast conditions, the study evaluated various filters. The application of a Laplacian filter for sharpening proved most effective. By emphasizing edges and structural features, the Laplacian filter allowed the CNN to achieve significantly higher accuracy compared to raw or blurred images.
Results and Discussion
The integrated system achieved a 95% prediction accuracy on a test set of 60 images (10 per species). The prototype successfully demonstrated the ability to trigger LED warnings and log data autonomously.
The "Sign as a Scientist" concept is perhaps the most significant contribution of this work. By transforming standard warning signs into distributed sensor nodes, government agencies can transition from relying on outdated historical data to real-time, data-driven ecological monitoring. The generated seasonal maps allow for the identification of specific "hot spots," enabling more targeted and cost-effective conservation efforts.
Conclusion and Future Work
The SWWS provides a scalable, solar-powered solution for reducing WVCs while simultaneously gathering vital ecological data. While the initial prototype is successful, future work will focus on extensive field testing to evaluate hardware durability in extreme weather. Additionally, expanding the species library and integrating the system with Vehicle-to-Everything (V2X) communication—sending alerts directly to vehicle dashboards—represents the next step in proactive road safety.
Summary updated