1. Introduction
Automated animal re-identification is crucial for ecological studies and wildlife conservation, but traditional solutions often require significant computational resources. There is a growing demand for cost-effective, low-power systems capable of performing re-identification in remote and challenging environments. This work addresses the problem of porting complex computer vision tasks like animal re-identification to energy-efficient microcontrollers. Models used in this study include optimized variants of MobileNetV2 and a custom-designed compact Convolutional Neural Network (CNN) architecture.
2. Related Work
Previous research on animal re-identification primarily focuses on improving accuracy using powerful GPUs, often overlooking constraints of embedded deployment. Concurrently, advancements in Edge AI have demonstrated the potential of running inference on low-power devices for simpler tasks like object detection or classification. This section reviews existing literature in both animal re-ID and embedded machine learning, identifying the gap our work aims to bridge by combining robust re-identification techniques with efficient embedded implementations.
3. Methodology
Our methodology involves several key steps to enable efficient animal re-identification on microcontrollers. First, we pre-process diverse animal image datasets to enhance feature extraction for re-identification tasks. Next, we design and train lightweight deep learning models, including a pruned MobileNetV2 and a custom TinyCNN, specifically for resource-constrained environments. These models undergo quantization and pruning techniques to minimize memory footprint and computational load. Finally, we deploy the optimized models onto selected microcontroller platforms (e.g., ESP32-CAM) and evaluate their performance using custom firmware.
4. Experimental Results
The experimental results demonstrate the viability of animal re-identification on microcontrollers, highlighting the performance of various lightweight models. We observed a direct correlation between model complexity and re-identification accuracy, with a trade-off in inference speed and memory usage. The optimized TinyCNN provided the best balance for typical microcontroller constraints, achieving acceptable accuracy with minimal resource consumption. The table below summarizes the key performance metrics across different models tested on an ESP32 microcontroller.
5. Discussion
The findings underscore the significant potential for deploying sophisticated computer vision applications like animal re-identification on edge devices. While accuracy may not match GPU-based systems, the achieved performance is sufficient for many practical wildlife monitoring scenarios. Future work will focus on further optimizing model architectures and exploring hardware acceleration capabilities of next-generation microcontrollers. The challenges of real-world environmental variability and power management also remain critical areas for continued research.