Research On Large-Scale Pedestrian Retrieval Algorithms Based On Deep Learning

The purpose of pedestrian retrieval is to find pedestrians with specific attributes in videos or images.Due to the development of deep learning and convolutional neural networks,pedestrian retrieval algorithms have made major breakthroughs in the past ten years.Pedestrian retrieval algorithms based on deep learning have been widely used in criminal investigations,intelligent surveillance and other fields.This thesis is dedicated to pedestrian retrieval algorithms with the use of deep learning.The main work can be summarized as follows:(1)Mutual Mean Teaching algorithm has achieved great success in unsupervised training for person re-identification,however,it requires two isomorphic models to work.This thesis introduces a new feature fusion mechanism that allows the algorithm work with two heterogeneous models.Experiments show that the new feature fusion mechanism has achieved better performance than original method,while the performance of heterogeneous models is limited by weaker models.(2)This thesis introduces a probabilistic ensemble learning algorithm for pedestrian attribute recognition based on meta-learning.This algorithm uses multiple basic classifiers and combines them with learnable weights via a meta learner.Experiments show that the performance of the proposed ensemble algorithm has been greatly improved,surpassing some classic algorithm in the field of pedestrian attribute recognition.(3)For pedestrian attribute recognition,this thesis tries to employ three ensemble learning algorithms,including equi-probability ensembling,non-equal-probability ensembling,and feature ensembling.Experiments show that the performance of non-equal-probability ensembling improves as the number of weak classifiers increases,and the performance is better than the other two ensemble learning algorithms.