Edge Deployment of Deep Networks for Visual Detection: A Review

With the development of artificial intelligence, there is an increasing demand for edge-computing visual sensors, particularly those integrated with visual object detection capability. However, the deployment of object detection models in edge computing environments faces technical challenges in many aspects, such as model size, inference speed, accuracy, and deployment optimization. This article systematically summarizes the knowledge related to the deployment of object detection at the edge-computing side, including mainstream models, deployment optimization methods, edge-computing deployment frameworks, and deployment devices, based on practical technical experience. The strengths and weaknesses of the various models and their applicability are analyzed, and the optimization and deployment of models on edge devices are explored in depth, with a focus on adapting them to the specific characteristics of edge-computing environments. This article is intended to provide valuable reference and insight to researchers and developers in the field, and therefore we provide deployment codes for reference at repository https://github.com/shouxieai/tensorRT_Pro.