Research On Object Detection Method Of Infrared And Visible Fusion In Complex Environment

Traffic target detection is an indispensable technology in the field of intelligent driving and intelligent transportation system.High reliability traffic target detection can provide real-time,reliable and accurate information for intelligent traffic and unmanned driving.At present,under the rapid development of automation and information technology to promote the traffic target recognition and detection in the ideal,structured environment have made outstanding achievements,but in all-weather conditions,especially the traffic environment,target identification in terms of recognition accuracy and robustness of sorts,combined with the traffic system as a system of safety first is related to the user’s life and property safety,the requirement of accuracy and robustness are more stringent.Therefore,it is necessary to study a object detection system with good real-time performance,high accuracy and strong expansibility for practical application..Focusing on practical problems,this paper studies two main aspects in detail:object detection based on multispectral information fusion and multi-scale object detection based on fusion recognition.Related research works are carried out based on convolutional neural network and image processing.The main research contents are as follows:First of all,this article on the basis of the development of artificial neural is introduced,according to the development order of neural network this paper introduces the principle and method in the field of visual identification and calculation of neural network characteristics,system structure and processing mode has carried on the detailed introduction,the back propagation algorithm,the error function in the field of target recognition and convolution calculation made mathematical explanation.The characteristics of deep convolutional neural network and a series of improved methods are emphatically introduced.This will pave the way for this paper to use this model for fusion detection.Secondly,in an all-weather environment,the optical camera will be affected by weather and illumination in the process of sensing data,resulting in problems at the source of information,leading to a large number of misdetection and missed detection after the subsequent target detection and recognition processing.This target detection model,which uses visible image as the only information source,makes the recognition effect of the recognizer strongly depend on the optical imaging conditions.In order to solve this problem,in this paper,the recognition model are introduced on the basis of the visible light and visible light image complementary,not influenced by optical conditions of infrared image information,studied in visible and infrared image fusion for object recognition methods,the two kinds of image feature level fusion way after the processing of basic network integration,performance optimization by introducing a multispectral information recognition,and using deep learning network learning multispectral image characteristics,reduce the effect of light on the recognition reliability,improve the recognizer recognition accuracy and robustness.Thirdly,in the actual application scenario,the physical scale difference of traffic targets is large,and the image data acquisition presents near large and far small,which makes the pixel area difference of different traffic objects in the whole image greatly difficult to identify,especially for the recognition of small-scale targets.In order to solve this problem,this paper on the basis of fusion recognition based on feature pyramid layers convolution neural network method are studied,on the multidimensional scale feature extraction,make full use of high dimensions much information on large scale target recognition performance and low dimension details for small scale target recognition performance is good,and through the rich in the process of training data set and multi-scale method raised the training model of multiscale target recognition accuracy and robustness,thus improve the identifier in the actual application scenario for multi-scale traffic target recognition accuracy and resolution.Finally,the combined application of information fusion and multi-scale recognition in traffic target recognition is discussed and analyzed,and the advantages and disadvantages of information fusion and multi-scale recognition on traffic target recognition are studied,which provides reference for further research.