Credit: Me, R. Y. Liao, Prof. J. T. Hsu, Prof. T. T. Lin, Dept of Biomechatronics Engineering, Dept of Animal Science and Technology, National Taiwan University, Experimental Dairy Farm
What I've Done:
- Architected Edge-Cloud ML System using Flask and AWS (IoT Core, S3, Athena, QuickSight); built CI/CD pipeline with GitHub Actions for over-the-air (OTA) software updates and IoT Core/MQTT for automated performance profiling.
- Drove resource-constrained deployment on Raspberry Pi with LiteRT/TFLite; quantified performance tradeoffs across inference optimization strategies, achieving 63% latency reduction and 40% RAM cut for real-time, low-power operation
- Delivered QuickSight reporting framework by collaborating with domain experts, saving 84% operational labor overhead
- Published research as 1-st author at top-tier bioengineering conference ISMAB 2024 (see publication page)
Motivation: AIoT for Livestock Monitoring
Monitoring behaviors in dairy cattle is crucial for efficient livestock
management and welfare enhancement, as variations in behavioral patterns often signal
health issues. Image-based behavior monitoring systems are non-invasive and suitable
for automated dairy farming practices, while these intelligent systems require substantial
computational resources to process consecutive video frames, which makes it
challenging to deploy behavioral recognition models on edge devices in dairy farms.
Our objective is to develop an efficient AIoT framework for image-based
behavior monitoring system for dairy calves, which can reliably recognize behaviors with affordable computational demands.
Building Efficient & Reliable AIoT Systems for Livestock Monitoring
Our AIoT system aims to improve livestock welfare and farm management efficiency by providing real-time behavioral analysis and health monitoring through computer vision and edge computing technologies. The system uses behavioral recognition algorithms to identify patterns in cattle movement, feeding behavior, and social interactions, providing farmers with immediate insights into animal health and wellbeing. By processing data at the edge and synchronizing with cloud services, our system is designed to help farmers make data-driven decisions and ensure optimal animal care through the power of AI and IoT.
System Overview
Based on the data gathered from our deployment at NTU's experimental farm, we have been able to show that our AIoT system achieves over 95% accuracy in behavior recognition tasks. Not only that, but we have seen that the system can process and analyze video streams in real-time with minimal latency, which is well above the performance requirements for practical farm applications! Combine that with the fact that over 90% of detected behavioral anomalies correlate with actual health issues, and our AIoT system has already been able to help farmers identify and address potential problems before they become serious.
The computer vision and machine learning techniques used within our AIoT system were developed in collaboration with the Department of Animal Science and Technology at National Taiwan University. This collaboration provided us with deep domain expertise in livestock behavior, animal welfare standards, and practical farm management needs. Our partners advised us on which behavioral patterns are most indicative of health issues, what types of monitoring data are most valuable for farmers, and how to maximize both system reliability and practical utility within the unique context of agricultural environments.
...And More!
We are still exploring ways technologies that help farmers take better care of their livestock, and researchers form better understanding of animal behavior and welfare. The systems has been deployed at NTU's experimental farm and has been scaled up to multi-device distributed systems. We're incredibly excited about the many directions we can take this project in to help more agricultural operations, and we know we've only just scratched the surface of what's possible with AI and IoT in agriculture. If you have any ideas or would like to collaborate, please contact us at polinnchen@gmail.com.
