Infrared Pedestrian Detection Based On Single Frame Multi-scale With Discernable Edge

Infrared pedestrian detection has become a hot topic in field of computer vision due to its important academic research and practical value,and has been widely used in vehicle-assisted,video surveillance,and military early warning.However,since the infrared image quality is usually poorer than the visible image,and pedestrian targets are complex,the design of high-efficiency infrared pedestrian detection algorithms still faces many problems.To improve the robustness of multi-scale infrared pedestrian detection of discernable Edge in single-frame images,especially the detection of low-infrared pedestrians in low-resolution images.Based on the analysis of characteristics of pedestrians,this dissertation firstly studies the methods of the maximum difference of neighborhoods to extracting a region of interest(ROI)region for reducing the generation of redundant candidate frames.On the one hand,The suspected head position search and multi-scale global template detection are conducted in the ROI region separately.On the other hand,the head template is weighted on the global template detection result to obtain a discernible contour single-frame multi-scale infrared pedestrian detection algorithm.The main research content is as follows.1.Review the present situation and existing problems in the field of infrared pedestrian detection.The principles and ideas of several infrared pedestrian character descriptors are given,and detection performance by simulating the results of each feature detection are compared.2.A method based on the maximum difference of neighborhoods of ROI is studied.Aiming at the problem that infrared pedestrian segmentation is easily interfered by background noise,considering that the target and the noise belong to the homogeneous and the non-homogenous region,the maximum difference threshold method of neighborhood is selected.Experimental results show that this method outperforms the BING algorithm in the number of ROIs extracted and the target accuracy.The accuracy rate can reach 96% in different backgrounds,and the positioning accuracy can reach 80%.At the same time,the positioning error and original accuracy can be maintained within 5% under salt and pepper noise and Gaussian noise.3.A method based on multi-scale infrared pedestrian head region auto-search is achieved.Aiming at the problem of diversity poses on infrared pedestrian,considering the stability of the head and the centrality of the edge direction,one eight-direction edge method is proposed.Establishing a reasonable head size scheme to achieve effective extraction of the head contour;On this basis,aiming at the problem of fewer pixels and missing details in the far-infrared pedestrian's head,a circular model segmentation strategy is constructed,and the brightness relationship between the sub-regions of each circle is judged according to the consistency of the cluster model.Experimental results show that this method performs well in multi-pose,occlusion and adhesion pedestrians.4.An infrared pedestrian detection algorithm based on improved features of global template detection and local template verification is Propose.Aiming at the problem of low robustness of infrared pedestrian detection at multiple scales,a weighted fusion classifier of global template and head template is constructed.In order to further enhance the ability of the global template,on the one hand,a texture-weighted HOI feature extraction method based on internal brightness features and residual detail features was studied for the near-infrared pedestrian.On the other hand,a entropy-weighted HLID feature extraction method based on the difference of pedestrian edge and background edge variation was studied for the far-infrared pedestrian.Finally,a new pedestrian head inspection feature template is selected based on the existing Haar_like feature template.The experimental results show that the proposed algorithm outperforms the DPM algorithm in detecting low-resolution far-infrared pedestrians,and the integrated detection rate of multi-scale infrared pedestrians can reach 85%.