Research And Implementation On The Method Of Pedestrian Detection Based On HOG-LBP

With the development of computer vision technology,pedestrian detection has received more and more attention and research.With the intelligentization of life and work,pedestrian detection,which is the core technology for applications such as assisted autopilot and intelligent video surveillance,its potential research value continues to increase.However,pedestrian detection is often caused by the variability of pedestrian attitude,occlusion and unfixed shooting angle,resulting in low pedestrian detection rate.Therefore,this thesis focuses on the detection of pedestrians occlusion.1)Aiming at the problem that pedestrians are prone to occlusion in traffic scenes,resulting in low pedestrian detection rate.This thesis develops the scheme of pedestrian upper body detection,designing part model,adding Edgelet feature to describe pedestrian local contour features,and learning 12 weak component model classifiers.Train the association between the sub-models of the mining components to improve the detection rate of the occluded pedestrians.By comparing experiments with other schemes such as HOG-LBP on different datasets,it is proved that the Part-Jointly model can improve the detection rate of occlusion pedestrians,and the integration function is determined by comparison experiments of different integration functions.2)The Part-Jointly model needs to perform feature extraction,classification,integration and other steps in the sliding window of the whole picture,which means that these steps need to be completely performed in the non-pedestrian area,resulting in large calculation and slow detection.Aiming at the above problems,this thesis develops a scheme for cascade detection of secondary network models.Firstly,the Pedestrian-RPN model is designed to filter the non-pedestrian area and output the pedestrian candidate box.The Part-Jointly model is secondarily detected on the pedestrian candidate frame and reduce the target area of the large Part-Jointly model,which speeds up detection and reduces detection time.By comparing the experiments with the single-layer network model on different data sets,it is proved that the performance index of the secondary network cascade detection model is almost unchanged,and the detection time is significantly shortened.3)In order to simulate the pedestrian detection function in the assisted driving system in the actual traffic scene,This thesis proposes to develop a pedestrian detection application on the mobile terminal Android smartphone,and use the mobile phone to shoot pedestrians on the traffic road,and simulate the pedestrian detection experiment.Considering the hardware resources,through the comparison of detection rate and detection time consumption,it is proved that the pedestrian detection application is basically available in the actual traffic scene,providing technical accumulation for assisted driving.