Design Of Deep Learning Based Infrared Machine Vision System

In recent years,the emergence of deep neural network has promoted the development of artificial intelligence technology.As the basis of many intelligent applications,visual perception system plays an important role in security monitoring,unmanned driving and intelligent robot.Most of the current visual perception systems are based on visible cameras,which give intelligent devices the ability to perceive the surrounding environment.With the increasing complexity of application scenes,the performance of visual perception systems based solely on visible sensors which will affected by light changes,can no longer meet the needs of practical applications.How to improve the object detection and measurement performance of visual perception system in complex environment and ensure its all-weather high reliability are the challenges that visual perception system must face in its further development.Due to its imaging characteristics,infrared sensors can feedback information in the case of complete darkness,and is robust to light changes and shadows,which can make up for the defects of visible sensors.Therefore,this thesis designed two sets of infrared visual perception embedded systems,one for body temperature measurement and the other one for intelligent driving.The main contribution are as follows:(1)In order to meet the requirements of rapid,high-precision and multi-person-time temperature monitoring in epidemic prevention and control,and to address the problems and challenges of the existing infrared thermometers with limited temperature measurement range and large maximum temperature measurement error(±0.5℃),a large-range and high-precision temperature measurement algorithm was developed.Through visible light image positioning the location of the face,thus obtaining the measurement results of infrared thermometer,and then use two layers of fully connected network to correct the result of the temperature measurement,and implements system based on binocular multispectral image.The experimental results show that the temperature measurement system designed in this thesis can reduce the maximum temperature measurement error from 1.73℃ to ±0.35℃ in the field of vision of the infrared thermometer,which verifies the effectiveness and feasibility of the system.(2)For automatic driving demand for all-weather,24 hours of work,and to address the problem that visual perception under the complex environment of night and fog is difficult,based on the imaging features of infrared image and the depth of ultra-high performance of neural network in the field of machine vision,by means of transfer learning a geared to the needs of the infrared target detection model of automatic driving,The prototype is implemented based on embedded system.The experimental results show that the system processes 720×480 infrared video images with the detection accuracy of 91.41%(m AP)and the reasoning speed of 29.65 FPS on the self-structured vehicle infrared data set.