Research On Head Re-identification Algorithm Of Bus Passengers Based On Deep Learning

Pedestrian re-identification refers to matching pedestrians at different places and time through computer vision technology.In the fields of intelligent security,shopping guide,and HCI,pedestrian re-identification technology has huge application potential.Compared with the method based on hand-crafted features,the method based on deep learning is end-to-end and can obtain more robust re-identification features.In recent years,researchers have obtained significant progress on many pedestrian re-identification datasets by deep learning technology.However,in crowded occasions such as buses and subways,severe occlusion makes it unrealistic to use passenger full-body images for re-identification.Therefore,the head area of passenger images is selected as an effective part to identify passengers.Therefore,in this paper,the head area of the passenger image is selected as an effective part to identify the passenger.However,due to the low resolution of passenger head images,few recognizable features,and high similarity between different samples,when recognizing passenger head using methods based deep learning,a large number of similar training samples makes the model fail to learn identifying features during training.In addition,because the heads of passengers photographed by the front and rear door cameras of the bus have different styles,matching the front and rear door passengers according to the head will affect the recognition performance of the re-identification model.This paper is based on deep learning technology,which mainly solves how to select effective training samples to improve the training effect of the re-identification model,how to generate passenger head images with different camera styles,and enrich the training sets,thereby guiding the re-identification model to better identify image under different camera.the main work of this paper includes the following:(1)Re-identification algorithm for bus passenger heads based on hard sample mining of triplet loss.Aiming at the effects of perspective,lighting,noise and other factors on the re-identification model,the training data is enhanced to improve the generalization performance of the model;Using the pedestrian re-identification model as a pre-training model to improve model training efficiency;Using different backbone networks to explore the impact on re-identification models;For the loss function,the experiment compare the effects of different thresholds and softening the loss boundary method on the re-identification model,and use different triple sample mining methods to verify the effectiveness of hard sample mining;On the basis of the triplet loss function,adding the constraint condition of camera labels to hard sample mining of triplet loss.The experiment results on the passenger head re-identification dataset show that compared with the benchmark algorithm,the average accuracy rate of the hard sample mining of triplet loss with the camera label constraint conditions is improved by 1.03%.(2)Re-identification method for bus passenger head based on CycleGAN generated picture.Aiming at the differences in camera styles between front and rear doors of buses,a camera style-transferred model based on CycleGAN is constructed.The style-transferred model was used to perform style transfer on the bus passenger head re-identification training set,and then the original image and the style-transferred image are used as training sets to train the re-identification model.Based on Tri Net,the positive effect of style-transferred images on rich dataset style was verified through experiments;Experiments based on other deep learning-based re-identificationmodels show that the generated style-transferred images greatly improve the re-identificationeffect of the model.