Study On Pedestrian Detection Based On Infrared Monocular

As an important problem in computer vision, pedestrian detection has long beenwidely concerned by researchers. In this thesis, we study deeply their sub-problems,which is pedestrian detection in monocular infrared vehicle scenes. Pedestriandetection is divided into regions of interest(ROIs) extraction and pedestrian detection.The main work of this paper is also carried out under these two aspects, the first is toextract pedestrian ROIs in monocular infrared vehicle scenes in order to avoid aexhaustive search to reduce the amount of computation; the second is to establish aninfrared pedestrian sample library, and the excellent visible light detection algorithmis extended to the infrared scene, combined with the ROIs extraction to achieve acomplete pedestrian detection system.Pedestrian ROIs extraction, aims to limit the search area as small as possible.Firstly, Otsu method is improved based on hot assumptions of infrared pedestrian andthreshold obtained from OTSU under statistical weighting function, in order to solveits problems which is pedestrians cannot be accurately divided. Then heat interferenceis filtered out by using the priori constraints, such as pedestrian aspect ratio, area. Last,ROIs is obtained by extending and merging of the suspected areas. Now, each ROIstill needs the pyramid exhaustive search because the size of pedestrians cannot bedetermined. Secondly, the possible pedestrian region is geometric constrained basedon the vehicle scene, and the upper and lower boundary constraints of the differentsizes of pedestrians are calculated out, and corresponding scaling factors aredetermined based on where the constraints of each ROI are, then ROI is smoothlydetected by detecting window after being scaled to appropriate size. Compared withtraditional methods, this method requires no additional detectors, so it can quickly andeffectively gets pedestrian ROIs, and greatly reduces the number of detectingwindows.The second part of the thesis mainly includes makingvehicle-infraredpedestriandataset, optimizing and adjusting the key parameters ofDeformable Part Model(DPM),andextending the method to the infrared scene, and theimplementation of a complete pedestrian detection system combined with ROIsextraction. As follows: vehicle-infrared Pedestrian Dataset is established under vehicle scenes with reference to visible dataset; the key parameters and componentweights of DPM, which includes the number of components, the number of parts,layers of twice pixels space, as well as the weights of parts, are optimized andadjusted. This thesis achieved a complete system combined with ROIs extraction.Firstly, scan area with pedestrian ROIs adaptive segmentation narrow detectionwindow. Secondly, determines scaling factor σ combined with geometric constraintsarea where each ROI is, and restricts the detection window to certain layers of thepyramid ROI. Also, when implementing the specific project, uses multiplying in thefrequency domain instead of spatial convolution to accelerate the computing.Experimental results show that, DPM has higher detection accuracy after optimized,and ROIs extraction brings about ten times improvement in the speed.