Pedestrian Detection And Re-identification Methods Research Based On Feature Extraction

In the widely used video surveillance system,the pedestrian re-identification task can be divided into two modules: pedestrian detection and pedestrian re-identification.These two modules have always been hot issues in computer vision applications.Pedestrian detection methods need to deal with a series of problems such as appearance variability,low pixel,small scale,humanoid object and occlusion.Most existing methods concentrate on neighboring feature extraction,but the scope of feature extraction needs to be improved.There are many challenges in the pedestrian re-identification methods,such as multi-view,changeable scale,occlusion and illumination.Existing methods can extract features such as color and texture,but these high-dimensional low-level features are unable to fully describe pedestrian individuals.In addition,the commonly used pedestrian features cannot handle the occlusion efficiently.The innovative research works on pedestrian detection and re-identification in this paper are as follows:The paper proposes a pedestrian detection method based on neighboring and non-neighboring features,main innovations are as follows: 1.The paper designs non-neighboring features based on structural symmetry and selects neighboring features with better balance in performance versus speed.Two types of features can be trained at the same stage;2.The paper introduce prior scale estimate in the mutual occlusion analysis algorithm to optimize the confidence scores of each body part.The experimental results on the relevant datasets show that the neighboring and non-neighboring features can efficiently identify human bodies with various scales and exclude humanoid objects.The improved mutual occlusion analysis algorithm can handle occlusion.Comparing with other models,this model has achieved better result in performance versus speed.The paper proposes a pedestrian re-identification method based on Gaussian distribution features.Main innovations of this method are as follows: 1.Fast super pixels preprocessing is applied to the input images,and the contour information is input as the attribute features.The Gaussian distribution features are used to model features with three levels and optimize the problem of excessive dimension;2.In the initial training stage,occlusion samples are automatically generated based on the occlusion sensitivity,and these samples are trained together with the original images in the retraining stage.Experimental results on multiple datasets show that Gaussian distribution features can establish complete pedestrian description.Super pixels processing,contour information input and adversarial occlusion samples training all increase the performance,and the performance of this method exceeds other single level feature models.