Research And System Implementation Of Pedestrian Re-identification Data Enhancement Algorithm

Pedestrian re-identification is a research hotspot in the field of computer vision,and has important application value in intelligent security and other fields.The current mainstream pedestrian re-recognition algorithms are mainly based on deep learning.As we all know,deep learning algorithms require a large number of training sets as a basis.Due to laborious and laborious manual labeling and privacy protection,large-scale pedestrian re-identification data sets are difficult to obtain,which limits the accuracy and wide application of pedestrian re-identification algorithms.Image processing algorithms such as image inversion,random cropping and erasure are used to overcome the problem of insufficient data sets,and up to a point it improved the ability of deep network learning network models to mine robust features on pedestrian data sets.However,the above-mentioned image enhancement algorithm is basically a simple transformation of the entire image,and cannot increase the feature information such as the pose and color distribution of pedestrians.In response to the above problems,this paper proposes new data enhancement methods in terms of pose feature and color distribution,which specifically include the following tasks:1)Process the posture information of the image based on the curve change feature of the mathematical trigonometric function.The algorithm is based on the discovery that trigonometric functions have natural advantages in pedestrian posture adjustment.Through in-depth exploration of the effect of trigonometric function curve changes applied to pedestrian posture processing,the image transformation can be adjusted from the original overall transformation to the transformation method of row pixel translation.,So that the pedestrian’s posture in the image can be adjusted in a variety of curves,and the pedestrian’s posture is changed.By randomly adjusting the parameters,a pedestrian sample can derive a variety of pedestrian postures,which greatly enriches the posture characteristics of the data set.2)Dynamically generate a new appearance of pedestrians based on the adaptive flood filling algorithm of boundary threshold.The algorithm can simply change the color of pedestrian clothing or background color,and can retain the basic color texture distribution characteristics of pedestrian samples.The algorithm is based on exploring the characteristics of the flood filling algorithm,and has developed a neighboring pixel gradient judgment algorithm,as an algorithm for judging the best color block,and works on pedestrian samples together,making the appearance of the new samples generated very realistic.This method can also deal with the problem of poor robustness caused by pedestrian occlusion.Because the new appearance generation method is random and local,the occlusion color block made by this method can be more advantageous than simply clearing the pixel block.Based on the research results of the above two works,a cross-camera pedestrian re-recognition system was developed to explore the application of pedestrian rerecognition in real life scenarios.