Research And Implementation Of Visual Object Detection Based On Collaborative Edge Computing

With the rapid development of computer and Artificial Intelligence(AI),lots of AI products have been applied in our daily life,especially face identification and self-driving.It is worth noting that nowadays self-driving is far from using in daily life.It's more common to use auxiliary-driving.Auxiliary-driving mainly include pedestrian detection,lane line detection,traffic light detection and so on[1].The key technology to realize functions above is object detection that can detect the location of object and classify,so it can help driver to avoid some danger.With the development of object detection through two-stage and one-stage detection,many detection model can achieve 59-155 FPS on TITAN X[2].But because of the computing power of common on-board device,detection model with high precision can't be deployed in those devices,which causes the conflict between real time and precision.Based on the idea of edge computing,this paper conducts some preliminary researches and explorations on the joint detection of mobile edge devices and road-side cloud and puts forward some ideas.The works include the following aspects:(1)Analyze the computing power of devices and train appropriate models for deployment.After analyzing the computing power of devices,the appropriate model is selected for training and quantizing.Finally the obtained model is deployed on the target devices.(2)Design and implement a mechanism for dynamic allocation of task.By sending heartbeat packet to get the feedback of network speed and computing power of devices and considering the cost function to determine which device the task is to be processed.This mechanism can ease the conflict between real time and precision effectively.(3)Data collection,preprocessing and experiment comparison.In the actual driving,the data is collected by cameras in vehicle and is processed into target format on road-side cloud.The object detection model is trained on the central cloud and updated through road-side cloud.Finally experiment comparison is designed to analyze the real time and precision of object detection platform based on collaborative computing edge.