TinyML-Driven Smart Real-Time Multi-Class Object Detection on Resource-Constrained Standalone Device

Resumen

Currently, the world is undergoing an AI and ML revolution, and we have seen tremendous growth in AI implementation across sectors over the last decade. Conventional AI-based systems were implemented in a cloud-centric environment, using servers with high processing power, ample storage, and high-speed internet, which consume significant power. This problem has driven the advancement of Edge-AI and TinyML, which have facilitated the implementation of intelligent visual systems on resource-limited embedded devices, enabling standalone device operation. In this paper, a standalone ESP32-CAM device is used to construct a real-time multi-class object detection system based on TinyML. For effective on-device inference, the suggested research strategy focuses on developing and using a quantized deep learning model based on the Edge Impulse architecture. To increase resilience across a range of environmental conditions, a bespoke dataset comprising four classes was produced and pre-processed. Using this method, we have produced an optimized machine learning model that is about 263 KB in size, with an inference latency of 133 ms and real-time performance of 7.5 FPS on the ESP32-CAM board. Under controlled circumstances, the results show an F1-score of 98.02%. In this approach, the end device itself runs an ML model and operates as a standalone device, offering strong data privacy, reduced communication, and very low latency. That’s why it balances accuracy, processing power, and real-time operation, making it best suited for small embedded devices.