Object Detection And Multi Label Classification For Edge Computing

With the advent of the intelligent era,in response to the national "new infrastructure" major strategic deployment,a large number of intelligent transportation,intelligent security monitoring and other artificial intelligence technology driven applications have emerged in the society.These applications often need to collect a large number of real-time image and video data,and require the use of intelligent algorithms for real-time and efficient image analysis and processing.However,the traditional cloud computing mode is difficult to cope with such harsh conditions because of limited network bandwidth,lack of real-time performance and poor security and privacy.Therefore,the best solution is to use the edge computing mode which can process and store data near the source side.However,there are still some problems in the edge platform,such as hardware performance,power sensitivity,limited computing power and so on,which directly affect the deployment of intelligent algorithms on edge devices.In view of the above pain points,this paper starts from two kinds of edge devices,the end side and the edge server,and discusses from two aspects of intelligent algorithm tuning and deployment acceleration,and realizes the algorithm cooperation on the two kinds of edge devices on this basis.The specific research contents of this paper are as follows:Firstly,the basic principles of image classification algorithm and object detection algorithm are introduced,and the basic situation of end side and edge server and the corresponding deployment technology are summarized.Secondly,according to the characteristics of rk3399 pro,an optimization method of network model compression is proposed for deployment algorithm.Through a series of migration training,sparse training,pruning,fine-tuning training and quantization,the amount of parameters and time-consuming of the algorithm model are greatly reduced while the target detection accuracy is not reduced too much.Then,for the end-to-side platform,the hardware advantages of NPU forward reasoning acceleration,RGA image processing acceleration and CPU multithreading acceleration are used to greatly improve the running speed of the end-to-side target detection algorithm.Then,according to the characteristics of edge server,for the bottleneck problem of feature fusion which limits the efficiency of reasoning in the algorithm,a lightweight improvement named grouping bilinear pooling is proposed.The improved multi label classification algorithm has the same classification accuracy as the original,and the amount of parameters and time consumption are greatly reduced.Then,openvino is used to deploy acceleration algorithm on CPU and tensorrt is used to deploy acceleration algorithm on GPU to give consideration to generality and efficiency.Finally,the advantages of the cooperation between the end side and the edge server are analyzed,and the data communication between the end side and the edge server is built,and the effect of the cooperation is tested and verified.Considering the characteristics of the edge platform,this paper discusses the optimization algorithm model,combines the development and deployment methods,fully exploits the performance advantages of the platform,and integrates the collaborative end-to-side object detection and multi label classification on the edge server,so as to make it an organic whole of good cooperation.