Making accurate object detection at the edge: review and new approach

摘要

With the development of Internet of Things (IoT), data are increasingly appearing at the edge of a network. Processing tasks at the network edge can effectively solve the problems of personal privacy leakage and server overloading. As a result, it has attracted a great deal of attention. A number of efficient convolutional neural network (CNN) models are proposed to do so. However, since they require much computing and memory resources, none of them can be deployed to such typical edge computing devices as Raspberry Pi 3B+ and 4B+ to meet the real-time requirements of user tasks. Considering that a traditional machine learning method can precisely locate an object with a highly acceptable calculation load, this work reviews state-of-the-art literature and then proposes a CNN with reduced input size for an object detection system that can be deployed in edge computing devices. It splits an object detection task into object positioning and classification. In particular, this work proposes a CNN model with 44 \(\times\) 44-pixel inputs instead of much more inputs, e.g., 224 \(\times\) 224-pixel in many existing methods, for edge computing devices with slow memory access and limited computing resources. Its overall performance has been verified via a facial expression detection system realized in Raspberry Pi 3B+ and 4B+. The work makes accurate object detection at the edge possible.