Design And Implementation Of CNN-based Vehicle Safety Driving Assistance System

With the improvement of people's living standards,automobiles have become an essential means of transportation for Chinese families.At the same time,with the advancement of technology,the ease of operation,safety and entertainment of automobiles have also been rapidly improved,enabling people to do more through cars.Things.In recent years,with the rapid development of artificial intelligence,computer vision is getting closer and closer to human recognition and judgment,and the intelligentization of automobiles is bound to be a trend.This article starts from computer vision and conducts research for the purpose of driving safety of automobiles.The main work has the following aspects:(1)Study and test the identification scheme of traffic signs.This paper first introduces the image recognition scheme based on statistical pattern recognition,which is difficult for the problem to be solved in this paper.Then introduce the artificial intelligence-based scheme.The most important solution of artificial intelligence is machine learning.By decomposing complex analysis problems into multi-layer network structures,and then using the powerful mathematical computing power of the computer,the large amount of sample data is continuously fitted.To derive better network parameters to solve complex analysis problems.The current mainstream machine learning network is CNN(Convolutional Neural Network),which plays an indispensable role in the field of image analysis.Then introduces a classic shallow neural network model LeNet-5.This paper uses this model to classify the traffic signs,and trains the LeNet-5 model through the Caffe machine learning framework tool.After 4000 iterations,it finally reaches 99.01%recognition accuracy.(2)Research and test road object detection schemes.Firstly,the common sliding window-based object detection scheme is introduced,and the Haar+ Adaboost classifier provided by OpenCV(Open Computer Vision)is used for verification.The face recognition speed is faster and the accuracy rate is higher,but the recall rate is lower because Many smaller faces are not recognized.Then introduce the object detection scheme based on the deep neural network MobileNet-SSD,which is more balanced in efficiency and accuracy,suitable for multi-target object recognition and classification,and then use the Caffe machine learning framework to perform target detection training on large-size pictures,which can reach 0.72.The average accuracy rate,and verified by actual road video samples,works well.(3)Research and test deploy complex neural network image recognition algorithms on low-cost embedded hardware.Firstly,the common low-cost embedded platform is introduced.The computational requirements for real-time applications of deep neural networks cannot be met.Then,an external neural network computing device with USB interface is introduced.The device can obtain very strong under very low power consumption.Floating point computing capabilities,and compatible with common Caffe and TensorFlow neural network models.In this paper,the MobileNet-SSD Caffe model is deployed to the Raspberry Pi 3 by the SDK provided by the neural computing stick.With the powerful FLOP capability of the neural computing stick,the 80 milliseconds reasoning time can be obtained,which can basically meet the requirements of quasi-real-time image recognition.