Research On Dynamic Target Recognition Method Of Fusion Of Visible Light And Far Infrared

Dynamic target recognition technology under video surveillance is an important subject in the field of computer vision and image processing.It is widely used in banks,shopping malls,hospitals and residential buildings.In recent years,many experts and scholars have analyzed and researched surveillance video,which has promoted the development of target recognition technology under video surveillance.This paper combines visible light and far infrared to identify moving targets under video surveillance,making the fused image more applicable and robust.The main research contents of this article are as follows:(1)In order to improve the accuracy and speed of infrared pedestrian detection and recognition algorithm,an improved three-frame difference algorithm is proposed in the dynamic target foreground segmentation,and the Sobel edge detection operator is integrated into the three-frame difference algorithm,which is free from the traditional algorithm detection.The foreground target has the disadvantages of “empty hole”,incomplete target contour and blurred edge information.In terms of dynamic target recognition,since the HOG feature performs a sliding window search through the full image to make the feature dimension higher,this paper changes the sliding window search strategy to perform a sliding window search on the ROI(Region of Interesting)region of the dynamic target foreground segmentation.The dimension of the HOG feature is reduced,and the background interference is also reduced,which improves the real-time and accuracy of the algorithm.(2)The traditional far-infrared and visible-light images have the disadvantages of information loss,poor contrast and no translation invariance.In this paper,the improved non-downsampled Contourlet transform method is used to decompose two images,and the high frequency sub-band adopts phase.The neighboring pixel difference is fused in a large manner,and the low-frequency sub-band is fused by combining the weighted average and the absolute value.The improved fusion algorithm has better performance than the traditional algorithm.(3)A target detection algorithm for the fusion of visible light and infrared images is proposed.First,feature extraction is performed on the two modal images,the feature images are fused during the extraction process,and the fused feature images are used in the RPN(Region Proposal Network).Preliminarily generate and classify the preselection box,then downsample the preselection boxes of different sizes to unify the images to the same size,and finally perform the precise positioning and classification of the preselection box to obtain the final detection result.In this paper,multiple groups of experiments are compared and analyzed,and the algorithm with the best detection effect is compared,and the feasibility of the algorithm is verified.