Lightweighted Real-Time Object Detection on a Custom Edge Device

With recent innovations in AI and software technology, on-device object detection has drawn significant attention. This technique enables real-time processing of visual data without the need for a connection to a distant server. However, deploying these models on resource-constrained edge devices presents several challenges. The primary obstacles stem from the limited processing power, memory, and storage capacity of these devices, as well as software issues. The current constraints make training artificial intelligence inefficient, as it requires substantial storage and computational power. Moreover, the development of devices based on ARM architecture demands the training and implementation of a customized model specifically designed for that edge device. This article discusses the development of a lightweight object recognition model that utilizes a TensorFlow Lite model and achieves a high accuracy rate of 94% on a custom edge device. This study also presents techniques for implementing this method using a custom file, demonstrates new performance metrics, and yields favorable results compared to existing benchmarks.