Research On Driver Smoking Recognition Algorithm Based On Convolution Neural Network And Implementation Of Embedded System

With the increase of car ownership,the corresponding traffic jams and traffic accidents are also increasing,traffic accidents caused by bad driving behaviors,such as smoking driving,are common,which do great harm to ourselves and others,many people have put car safety in the first place.In recent years,the computer hardware is constantly updated,and deep learning neural network has been developed rapidly,which is widely used in face recognition,driverless and other fields.This paper focuses on the field of machine vision,uses the deep learning convolution neural network algorithm to identify the driver's smoking state in real time,and transplants the algorithm to the embedded computing platform to complete the real-time monitoring task of the driver's smoking status in the embedded terminal.Considering that the parameters of the classical network algorithm are too large,and the computing power of the embedded platform is limited,this paper adopts the lightweight network structure MobileNet,draws lessons from the Inception structure to extend the receptive field of feature extraction,puts forward lightweight classification network structure MobileNet-Inception.Based on the Caffe deep learning framework,the training and testing of MobileNet-Inception algorithm is realized by DIGITS tool.Finally,the trained model is transplanted to TensorRT reasoner,deployed to Jetson TX2 platform to realize the task of real-time monitoring the smoking status of drivers in the embedded terminal.In order to comprehensively consider the running effect of the algorithm deployed in the embedded terminal,this paper has carried out experimental tests from many aspects.At the host side,the classical algorithm AlexNet,GoogleNet and lightweight algorithm MobileNet are compared.The results show that the accuracy of this algorithm is 4% higher than that of mobilenet,the memory usage increased by only 2.7MB,and the performance of this algorithm is improved while maintaining the characteristics of lightweight,so it is suitable for embedded platform.The embedded terminal is tested in three aspects: The implementation efficiency of TensorRT and Caffe frameworks is compared;the execution effect of the algorithm in different scenarios is comprehensively tested;the efficiency of the classical algorithm AlexNet,GoogleNet and lightweight algorithm MobileNet is compared.The results show that the algorithm and deployment method in this paper make a trade-off between accuracy and speed,meet the task requirements in terms of robustness andenergy consumption,and can effectively identify.