Lightweight Real-time Object Detection System Based on Embedded AI Development Kit

Abstract

Lightweight object detection focuses on the lightweight and real-time performance of the model, aiming to reduce the size of the model as much as possible without significantly reducing the detection accuracy, so that it can be deployed in practical application scenarios. This system improves the MobileNet-SSD object detection algorithm and uses standard datasets for training and testing. Through channel pruning, the parameters of the model are greatly reduced while the accuracy remains unchanged, and the model compression ratio is 1.17:1, which reduces the model’s occupation of the device memory capacity, and obtains about 44% improvement in detection speed. Finally, the trained model is deployed on the embedded artificial intelligence development kit EAIDK-610 to process and detect the collected video content in real time. This system can be extended to practical detection tasks with limited computing resources in various specific occasions.