Research On Forward Vehicle Detection And Taillight Status Judgment Algorithm Based On Deep Learning

Intelligent driving is a vehicle that changes the driving strategy according to various sensor devices,perceives the surrounding environment and its own state,and automatically makes decisions based on the surrounding environment and its own state without driver intervention.Intelligent driving can avoid unnecessary driving errors of many drivers,reduce the occurrence of trafic accidents,improve the utilization of traffic resources,and reduce people’s travel costs.Therefore,intelligent driving technology has important research significance.In recent years,with the continuous research of deep learning technology in the field of computer vision,computers have surpassed humans in some visual tasks.In intelligent driving,compared with other sensors such as radar,ultrasonic,etc.,the image sensor can obtain more information,and is also the main way of information acquisition by the driver during driving.The main work of the thesis is as follows:Firstly,the research background and significance of intelligent driving are expounded,and relevant theoretical research on image processing algorithms is done.The image preprocessing technique is studied.The differences,advantages and disadvantages between the traditional image processing algorithm and the depth learning based image processing algorithm are compared and analyzed.The algorithm using deep learning is determined as the image processing algorithm.Secondly,a depth detection-based target detection algorithm is designed and implemented.Based on the convolutional neural network,the algorithm is designed in detail.The network is based on the idea of regression and uses high-level features with low resolution.The characteristics of high semantic information and low-level features have the characteristics of high resolution and low semantic information.The feature fusion of high-level and low-level features can improve the adaptation of the network to scale changes and the detection of small targets.Compared with the detection framework of HOG+SVM,the detection algorithm of this paper has an accuracy of 95.70%and a recall rate of 96.80%under the threshold of IOU>0.5.The experimental results prove the superiority of the proposed algorithm.Then,for the classification task of the taillight state of this paper,the data annotation tool is developed to solve the annotation of the classification data of the vehicle lamp.The classifier based on convolutional neural network is designed and designed,and the single-label network model and multi-label are designed respectively.The network model of the tag;in order to increase the richness of the sample data the data is expanded and enhanced in different ways.The experimental results show that the multi-label classifier has certain advantages,and the classification accuracy of all types has reached more than 95%.Finally,based on the above implemented algorithm,the intelligent driving subsystem—the rear detection tracking and the taillight status judgment system is designed and developed.The system can be input with a real-time camera or captured video,with front tail detection tracking and taillight status classification.The experiment proves that the system can detect and track the vehicle in front well,and can accurately determine the state of the taillight.Real-time requirements can be achieved in terms of performance.This paper studies the key technologies of visual perception in intelligent driving,and makes a useful attempt on this basis.The research results have certain practical application value,and have certain reference value for the application of intelligent driving visual perception technology.