The endangered native cat species of Taiwan, the clouded leopard, has received significant attention in recent years. With the rise of urban and rural development, threats to the environment and the lives of wildlife are increasing. Currently, the number of wild clouded leopards has fallen to fewer than 500. Statistics show that between 2015 and 2018, there was at least one roadkill of a clouded leopard per month. To address this, Taiwanese startup DT42 (Zhuzhuo Technology) is collaborating with government and academic institutions using an artificial intelligence image recognition system to protect local animals through more timely and efficient methods, fostering an environment where humans and animals can coexist harmoniously.
The roadkill warning system, developed by the Ministry of Transportation and Communications' Highway Administration, the Center for Endemic Species Research and Conservation, and the Department of Mechanical Engineering at National Chung Hsing University, in collaboration with DT42, has been operational since May of this year in a demonstration area on Provincial Highway 5 in Zhuolan, Miaoli. The system detects whether vehicles traveling on the road are speeding too high, and simultaneously activates a built-in animal recognition system and an acoustic and light emission system to alert leopards currently on the road or preparing to cross it, ensuring they stay off the road, thereby achieving the intended goal of reducing roadkill.
This system is not only designed for the endangered clouded leopard, but also hopes to protect animals such as the white-nosed leopard and ferret badger that also live on this road section. It can even reduce the unfortunate incidents of wild dogs and wild cats being hit by vehicles that pass by because they do not have time to dodge.
Challenge: The data required for animal identification training is vast and complex
The technology and infrastructure required for a roadkill warning system are not yet fully developed, and the initial data collection and configuration work is tedious and complex. The team uses AWS cloud computing to identify animals, but considering the bandwidth required for real-time identification and data transmission, and the overall operational response speed being slower than expected, it is necessary to further introduce analysis and calculation functions on the terminal devices.
DT42 Project Manager and Co-founder Chung Wan-ka said, "Because vehicles travel at very high speeds, the roadkill warning system is activated immediately if the system detects a vehicle exceeding 40 kilometers per hour. Animals are erratic, so the generation of light and high-frequency sound is necessary to keep them away from the road and prevent them from being hit by vehicles. If the roadkill warning system can be activated more quickly, it will be more effective and save animal lives."
Solution: GPU acceleration improves the learning efficiency of the recognition system, allowing the roadkill warning system to operate smoothly.
To accelerate data analysis and organization, and to reduce the time and labor required for subsequent image data labeling and correction, the design team chose to integrate NVIDIA's Jetson TX2 platform, designed for edge devices. Its lightweight, power-efficient solution integrates CUDA and CuDNN computing architecture technologies to improve training efficiency, and uses TensorRT for inference acceleration. This allows a learning model to be completed in just three hours and quickly deployed on edge computing devices, ensuring the smooth operation of the roadkill warning system.
The Jetson TX2 platform enables terminal devices to complete preliminary analysis with higher-performance computing, without relying entirely on cloud-based collaborative computing resources, and will not be affected even when the device is offline.
With the help of artificial intelligence technology, the system can reduce roadkill through image recognition and other systems such as vehicle speed detection. Through GPU-accelerated image recognition, it can quickly determine whether there are animals preparing to cross the road from the image details, allowing the roadkill warning system to function normally.
The Endemic Species Conservation Research Center team also installed fences in the demonstration area to prevent animals from crossing the road, and cleared culverts that animals could walk through as an aid. By guiding animals to reach the other end of the road by other means, the rate of roadkill is reduced.
The actual installation of the roadkill warning system is mainly to cooperate with existing buildings and will not damage the environment or affect the original landscape. However, Dr. Zhang Junwei of National Chung Hsing University is still concerned that individuals may steal or damage the equipment, or use it to determine the possible locations of protected animals, thereby causing poaching and other problems. Therefore, the system facilities will be further disguised during the construction process, and the actual installation location of such facilities will not be disclosed to the public.
Impact – It is expected that artificial intelligence will be used to implement more animal conservation in the future
After the roadkill warning system was introduced on the demonstration road section, the number of roadkill accidents involving tigers was significantly improved, with only one accident occurring within three months. The Ministry of Transportation is very satisfied with the results of this cross-border cooperation program and expects to promote such applications to more sections of roads prone to roadkill in the future.
Assistant Professor Ya-yu Chiang of the Department of Mechanical Engineering at National Chung Hsing University stated, "The test results of the clouded leopard roadkill warning system are encouraging. We are actively collaborating with the government to replicate this successful experience in conservation programs for related animals, and to promote it to other parts of Taiwan and globally, contributing to ecological conservation through AI."
Animal conservation has been a global focus in recent years, with governments investing significant resources and initiatives. Many major technology companies are also leveraging innovative technologies to provide more effective solutions. DT42 believes that the potential for future development in animal conservation lies in the application of artificial intelligence (AI). Therefore, by working from the grassroots level and collaborating with local farmers to identify fundamental issues and address pain points in a manner tailored to their needs, we can achieve more efficient conservation outcomes without impacting existing ecosystems or local development.
