Design And Implementation Of The Team Collaborative Safety System Based On Visual Deep Learning And Vehicle Communication

With the continuous development of the transportation industry,more and more attention is paid to the safety of road driving by detecting moving targets such as vehicles and pedestrians on the road.In recent years,with the development of deep learning,tasks of computer vision have been replaced by deep learning.And deep learning is also used to solve target detection problem.At the same time,because the mobile device is a safe and private computing resource which does not need to be connected to the network,the target detection based on deep learning technology has become a hot topic in the current study.In this paper,firstly we introduce some common methods of target detection algorithm,and compare advantages and disadvantages of various methods.Then we introduce our design of deep learning framework based on mobile devices,which consists of parallel processing layer,control layer and application layer.And we implement convolution layer,max pooling layer,fully connected layer and activation function based on the OpenCL,a parallel and multi-core processing framework.We propose and implement the optimized acceleration method based on OpenCL memory model.We preload the convolution input and convolution kernel into local memory and constant region from global memory,and use branching avoidance method to reduce the waiting time of work items.A map-reduce model is used to divide the convolution layer and the pool layer,and the data of the adjacent work node are exchanged after each layer is calculated.At the same time,we design and implement the task redistribution mechanism based on exponential smoothing in order to solve the cask effect due to the unequal computing power of different work nodes in the distributed system.The calculation ability of each work node is predicted using exponential smoothing,and computing tasks are allocated according to different calculation ability,thus improving the overall computing acceleration ratio.Finally,we test the performance of our system by two to six mobile devices,and then compare the execution delay and acceleration ratio between averaging method and prediction method.Experiments show that the prediction method has higher performance than averaging method,and has 1.95x to 5.94x acceleration by 2-6 work nodes,while the averaging method has 1.69x-5.78x acceleration.Both of them have the same overall acceleration performance of 3.35x and 3.36x by 6 work nodes.