Low Power Artificial Intelligence on Edge Devices: Implementation on Raspberry Pi and ESP32, an Ultra-Low Power Microcontroller for Edge AI Applications
Keywords:
Edge AI, ESP32, Internet of Things (IoT), Low-power artificial intelligence, Machine learning on edge devices, QoS, Raspberry Pi, TensorFlow Lite, Ultra-low-power microcontrollerAbstract
Low-power AI on edge devices is something that is getting a lot of attention lately. It lets systems handle data right where it is, on the device itself, instead of always sending everything to the cloud. That cuts down on delays, keeps things more private, and uses less network and power overall. That makes it pretty useful for health monitoring or smart homes, where quick responses are required. The goal of this study is to build and test simple AI models that perform well on limited hardware, without sacrificing much accuracy or slowing down. Hardware selection is the initial step, based on processing capability, memory capacity, power consumption, and connectivity requirements. Raspberry Pi is good for heavier tasks, like spotting objects in images or video stuff, since it has more power and RAM, plus it runs libraries easily.
On the other hand, ESP32 is for lighter jobs, say watching sensors or recognizing voice commands, because it consumes less power and has built-in WiFi and Bluetooth. For implementation, data are collected from cameras or sensors, depending on the application. The data are then preprocessed by cleaning, resizing, and splitting them into training and testing datasets. This process improves the performance and generalization capability of the models. The models are built in Python with TensorFlow Lite, using things like basic CNNs or TinyML that do not need many resources. To improve computational efficiency, optimization techniques such as quantization and pruning are applied to reduce model size and computational requirements. Deployment strategies vary depending on the target hardware platform and application requirements. On Raspberry Pi, OpenCV and TensorFlow are used for images, while ESP32 uses Arduino or Embedded C for simpler embedded work. Evaluating means checking accuracy, how long inference takes, CPU and memory use, and power draw. Results show Raspberry Pi handles complex tasks with higher accuracy, but ESP32 shines when power and portability matter more, like for always-on devices. Both platforms work for edge AI; it depends on the app and power available. Some applications might need one over the other, and that choice is not always straightforward.
References
A. Moursi, U. Qidwai, S. R. Hasan, A. A. Ali and A. Erradi, “AI-based hardware Trojan intrusion detection system for IoT and edge devices,” in IEEE Access, vol. 13, pp. 191689–191706, 2025.
A. Alenezi and Y. Alabaiadly, “Digital twins and Edge AI for small/off-grid water systems: A review,” in IEEE Access, vol. 14, pp. 39960–39978, 2026.
C. -H. Chen, C. -H. Chang, C. -W. Peng and G. -W. Lan, “Enabling AI-driven real-time communication and computation with DDS for IoT-assisted home care,” in IEEE Internet of Things Magazine, vol. 9, no. 2, pp. 41–48, Mar. 2026.
N. Faris Ali, H. Al-Nashash, I. M. Elfadel and M. Atef, “A review of photoplethysmography-based blood pressure monitoring: From cloud-based machine learning to TinyML edge deployment,” in IEEE Access, vol. 13, pp. 204210–204240, 2025.
C. Yahyati et al., “A systematic review of state-of-the-art TinyML applications in healthcare, education, and transportation,” in IEEE Access, vol. 13, pp. 204513–204562, 2025.
V. Ieropoulos, E. Anthi, T. Spyridopoulos, P. Burnap, P. Carnelli and A. Khan, “Federated detection at the edge: Collaborative anomaly detection for resource-limited IoT,” in IEEE Internet of Things Journal, vol. 13, no. 11, pp. 24205–24218, Jun. 2026.
C. N. da Silva and C. V. S. Prazeres, “Tiny federated learning for constrained sensors: A systematic literature review,” in IEEE Sensors Reviews, vol. 2, pp. 17–31, 2025.
M. Akhi, C. Eising and L. L. Dhirani, “Securing IoT using lightweight TCN for edge deployment on Raspberry Pi 4,” in IEEE Open Journal of the Communications Society, vol. 7, pp. 442–460, 2026.
A. Abdelli et al., “Field validated hybrid ESP-NOW and long range IoT monitoring system for energy autonomous precision agriculture,” in IEEE Access, vol. 13, pp. 207511–207526, 2025.
R. de la Fuente, L. Radrigan and A. S. Morales, “Enhancing predictive maintenance in mining mobile machinery through a hierarchical inference network,” in IEEE Access, vol. 13, pp. 59480–59504, 2025.
A. Cerioli et al., “Efficient detection of microplastics on edge devices with tailored compiler for TinyML applications,” in IEEE Access, vol. 13, pp. 90970–90982, 2025.
I. Lamaakal et al., “A comprehensive survey on tiny machine learning for human behavior analysis,” in IEEE Internet of Things Journal, vol. 12, no. 16, pp. 32419–32443, Aug. 2025.
M. J. C. S. Reis, “Scalable intrusion detection in IoT networks via property testing and federated Edge AI,” in IEEE Access, vol. 13, pp. 153244–153262, 2025.
A. Nawaz, H. H. M. Ramzan, X. Yu, Z. Zou and T. Westerlund, “Blockchain powered edge intelligence for U-healthcare in privacy critical and time sensitive environment,” in IEEE Journal of Biomedical and Health Informatics, pp. 1–13. 2025.
R. Gholami, H. D. Jahromi and S. Sedaghat, “Design and implementation of a near real-time human detection robot using YOLO framework and IoT technologies,” in IEEE Access, vol. 13, pp. 175960–175983, 2025.
O. B. Eldine, M. Selim, N. El-Makky and N. M. Ghanem, “TinyML prototype of infant cry classification system,” in IEEE Access, vol. 13, pp. 147604–147614, 2025.
