Research And Application On Edge Computing Architecture For Portable Devices

Portable devices are small devices that can be carried,disassembled,handheld,small in size and low in power consumption for assisting work and life,such as voice helper robots,guide robots,etc.These devices have real-time human-computer interaction,built-in sensors,and certain visual and voice interaction capabilities(such as a portable guide robot).Deep learning can solve the problem of visual and voice interaction very well,but running a deep neural network model requires a huge computational overhead,which is a challenge for the design of portable devices with high real-time requirements.The real-time performance of the scheme for data transfer to the cloud is limited to the network and cannot be guaranteed if the network environment is poor.Edge computing emphasizes the migration of computing tasks and decisions from the cloud to the network edge to improve the real-time performance of portable device systems.This paper presents an edge computing architecture PEN(Phone+Embedded board+Neural compute device)for portable devices.The architecture uses embedded single boards to build device bodies,decision centers on smartphones,and devices that can be used for deep neural network inference to speed up image or voice processing.PEN abstracts the hardware and software functions of portable devices in a "service-oriented" way.Smartphone clients can integrate these services and build decision centers through remote calls using visual programming tools or programming languages.In order to verify the universality of this design framework,this paper designs a set of portable device-take-out assistant which can be used to assist the delivery of takeaway based on PEN architecture.This device integrates sensors and motion devices such as camera,steering gear and ultrasonic module,and runs the neural network model through the Ganzhi K210 chip to complete the target detection task.In this device,Edge Server provides multiple hardware function services that can be invoked by smartphone clients.The mobile side implements business logic on the decision side and acts as a gateway for data cleanup,filtering encapsulated data and forwarding it to the cloud.The last part of this paper is the test and data comparison of the actual scenario to prove the usability of PEN.