Research On Night Vehicle Detection And Recognition Technology Based On Image Processing

With the rapid development of economy and science and technology,driverless technology is gradually rising.The existing driverless vehicles mainly use infrared,radar and other sensors to sense the external environment to make positive coping strategies to ensure the safety of driverless vehicles.The high price of infrared,radar and other sensor equipment with superior performance seriously restricts the popularization and wide application of driverless vehicles on the road.As a new perception technology,the vehicle recognition system based on machine vision is not only simple to install and cheap,but also can obtain abundant information of road,vehicle and other external environment in front of it.Therefore,the vehicle detection based on machine vision has become a research hotspot of driverless direction.In this paper,the problem of poor image detection accuracy in the case of insufficient illumination at night environment is studied.The low illumination image enhancement and night road target detection technology are studied.The purpose is to meet the needs of vehicle detection at night visual effect and night image processing real-time,but also to improve the real-time and accuracy of vehicle recognition.The main contents of this paper are as follows:Firstly,the development status of low illumination image enhancement and night road target detection at home and abroad is analyzed by consulting the relevant literature.At present,although the low illumination image enhancement technology has a certain effect on image enhancement,it will lead to excessive image enhancement,image distortion and other problems.Moreover,the existing road target detection technology at night has some problems,such as limited application environment,low detection accuracy,difficult detection model embedding and so on.In order to solve the above problems,the research ideas of low illumination image enhancement and night road target detection technology are determined.Secondly,a data acquisition scheme of low illumination road vehicle images is designed,which collects road images in different scenes at night.Combined with bdd100k public database of automatic driving in deep learning field,and through data enhancement technology,a total of 5647 images are obtained,and the road vehicle recognition data set at night is constructed.The labeling method is determined,and the data set is labeled.The collected sample data basically represents the characteristics of vehicles at home and abroad,the diversity and complexity of the environment,and reflects the real environment of daily driving at night.Then,the advantages and disadvantages of traditional image processing algorithms for low illumination image processing are analyzed.Traditional image processing algorithms can improve the brightness effect of low illumination images,but there are also problems such as image quality degradation and image distortion after enhancement.In order to improve the brightness of the image and ensure the naturalness of the image,this paper proposes an improved image enhancement algorithm based on the traditional Retinex image enhancement algorithm.The algorithm realizes the enhancement by calculating and estimating the illumination map of the low illumination image,which is different from the traditional Retinex algorithm,which decomposes the image into illumination image and reflection image,but only estimates one illumination variable,The experimental results show that the algorithm has the highest score in information entropy,gray variance product and edge intensity,and also has advantages in real-time analysis.Finally,aiming at the low illumination image vehicle recognition technology,the yolov4mobilenetv3 neural network model is proposed for vehicle recognition and detection.The main feature extraction network of the detection model is mobilenetv3,which reduces the number of parameters of the network model,improves the embeddedness of the whole model,and improves other modules of the detection model.Through the experimental comparison and analysis,compared with the general detection and recognition model,the improved neural network model improves the embeddedness of the whole detection model on the basis of ensuring the recognition accuracy,and the detection effect is better.