| Summary: | AIoT is a novel concept that focuses on combining the intelligence of AI applications with the connectivity provided by the traditional IoT infrastructures. However, the high computational complexity of deep convolutional neural networks (DCNN) curbs their deployment on IoT devices with limited computational resources. This report presents the implementation of an AIoT system and the development of AIoT applications. Quantization techniques were utilized to compress and accelerate a real-time object detection model on a typical resource-constrained IoT device. IoT communications and backend services were set up to support smooth information networking between the edge devices and user client. Demo applications that allow users to remotely access the object detection result yielded on edge were presented in this report. The established platform unlocks abounding potential user scenarios, e.g. escape room game. The research regarding the quantization techniques for neural networks in AIoT applications was also conducted. Mix-bitwidth networks searched out by the proposed method achieved competitive accuracy with fewer parameters and computational footprint than the uniform bitwidth counterpart. which is promising for the efficient deployment of DCNN on resource-constrained edge devices in the future. This work on quantization is prepared to be submitted to the conference of ICARCV 2020 for review.
|
|---|