Two-dimensional Fake Pedestrian Recognition Based On Light Field Camera

As a branch of object detection,pedestrian recognition has been widely used in intelligent transportation,monitoring system and image retrieval due to its high efficiency and high accuracy.Although pedestrian recognition technology has achieved fruitful research results in the past decade,the complexity of the scene and the diversity of pedestrian postures and gestures still pose a serious challenge to the pedestrian recognition.At the same time,the traditional camera or video acquisition equipment only records the two-dimensional information of the scene and loses the three-dimensional information of the scene,which makes a large number of two-dimensional false pedestrian images in the daily life scene,and easily leads to the occurrence of pedestrian misidentification.Based on the fact that the optical field camera can record the depth information of the scene,this paper proposes a scheme combining the light field camera with the efficient pedestrian recognition algorithm,which provides a new idea for solving the problem of two-dimensional fake pedestrian recognition.The main contents of this paper are as follows:1.Firstly,some classical INRIA dataset,TUD dataset,VOC2007 dataset and Caltech dataset are taken as samples for pedestrian recognition experiments,which proves that these datasets have the limitation of lacking scene stereo information.Secondly,the first dataset of 1091 pedestrian light field images including positive and negative samples was set up by using Lytro-Illum light field camera.Positive samples include pedestrian images with different postures and dresses in different scenes;negative samples include non-pedestrian images such as pedestrian traffic signs,LCD pedestrian images and human posters.Finally,the official Lytro Desktop software is used to post-process and improve the pedestrian light field image dataset,and the two-dimensional image and depth image corresponding to the light field image are obtained.2.Firstly,some classical INRIA dataset,TUD dataset,VOC2007 dataset and Caltech dataset are taken as samples to carry out pedestrian recognition experiments,which proves that these datasets have the limitation of lacking scene stereo information.Secondly,a dataset of 1091 pedestrian light field images including positive and negative samples was established by using Lytro-Illum light field camera.Each light field image corresponds to a two-dimensional image and a depth image.The positive samples of the dataset include pedestrian images with different postures and dresses in different scenarios;the negative samples include pedestrian traffic signs,LCD pedestrian images and non-pedestrian images such as human posters.Finally,the official Lytro Desktop software is used to post-process and improve the pedestrian light field image dataset.3.Another two-dimensional fake pedestrian recognition scheme based on local binary patterns(LBP)feature extraction algorithm and support vector machine(SVM)classification algorithm is proposed,and the selection of SVM kernel function is experimentally studied.The results show that the SVM pedestrian recognition model based on arbitrary kernel function with small samples still has high recognition accuracy.which proves the stability,advancement and portability.of this recognition scheme.