Traffic Object Detection Based On Multispectral Image Fusion Under Variable Illumination Conditions

With the rapid development of computer vision field,traffic object detection and recognition based on machine vision has gradually become a research hotspot.Nearly a dozen years,object detection technology has made great progress from the initial traditional manual feature extraction of features to the current mainstream deep learning method,but due to the complexity and changeability of the traffic environment,the object detection algorithm is still restricted by environmental conditions such as light and temperature,and it is still difficult to accurately and quickly achieve the object detection in the traffic environment.This paper focuses on the detection of traffic objects under the environmental influences of lighting changes,weather and so on.The main research contents are as follows:(1)In view of the fact that visible light images are susceptible to environmental factors such as illumination and weather,single-mode images are not robust in object detection,this paper explores the dual-mode fusion enhancement method of infrared and visible images,and the dual-mode information can complement each other and provides rich contour and detail information for accurate object localization and classification.We introduce a nonsubsampled contourlet image fusion algorithm,and design fusion rules for high and low frequency information to obtain fused images.In the terms of public datasets,the proposed method is superior to other fusion algorithms under multiple evaluation indicators,and compared with single-mode object detection,the fusion image has a greater improvement in detection accuracy.(2)Under the complex traffic environment,the key object recognition is vulnerable to background noise,and object missed detection is relatively high.Thermal objects such as traffic vehicles and pedestrians have significant characteristics in infrared imaging,in order to quickly and accurately extract thermal objects in infrared images,this paper analyzes the salient area of infrared thermal and introduces the salient area of deep learning area detection algorithm,through multi-scale transform information fusion method for improvement of the salient object detection,thereby enhancing the objects information in the fused image.On public datasets,this method can achieve better multi-scale saliency detection results on infrared images.(3)In order to increase the attention to the image objects targets,this paper designs a perception model to obtain the light sensitivity of the fusion images,and calculates the illuminance weights of visible-infrared fusion images based on gray scale indicators and gating functions and convolutional networks,and realizes the adaptive weight secondary fusion of infrared saliency map and fused image,and balances the image background and salient object information.After experimental analysis,traffic objects detection under the complex fusion mode has a better effect.(4)On the basis of the above image processing,an object detection framework of image fusion and salient attention mechanism is proposed,and an object detection datasets of secondary fusion images is prepared,on which experimental evaluation and analysis of complex fusion images are carried out.Finally,traffic object detection system of multispectral image fusion under variable illumination conditions is designed,and the final object detection result is obtained through the multi-modal image preprocessing module and the object detection module.