Feature Diversity Achieving For Deep Learning Based On Person Re-identification

Person re-identification(Re ID)is a technology that uses computer vision technique to identify whether there is a target person in images or videos across different cameras.In recent years,deep learning has achieved great success in many computer vision tasks,person Re ID based on deep learning is widely used in public safety,intelligent monitoring,and smart city,etc.For the person Re ID,the critical problem is how to extract robust and discriminative features.Based on deep learning and its application in person Re ID,this thesis dedicates to the development of algorithms in the field of person Re ID.The main contributions are as follows:(1)Based on part-based convolutional baseline(PCB),we propose a weighted PCB algorithm,which combines the global feature and local part-based features in a weighted form.Experiments show that the proposed weighted features are more discriminative and the proposed algorithm performs better than the original PCB.(2)For person Re ID,most state-of-the-art deep models depend heavily on Image Net pre-training.The Omni-Scale Network(OSNet),as a lightweight model,was recently proposed in achieving excellent training-from-scratch performance for person Re ID.We propose to construct a more powerful model,termed PFH-OSNet,based on the idea of learning pyramid feature hierarchy over OSNet from scratch.With a number of cooperative lateral branches,PFH-OSNet shows faster improvement in m AP performance compared to OSNet when trained from scratch.Empirical evidence demonstrates that PFH-OSNet achieves state-of-the-art performance on popular person Re ID datasets,including Market1501,Duke MTMC-Re ID and CUHK03.(3)Although the performance of person Re ID has been greatly improved in recent years,existing algorithms still have poor capability on generalization,which make them very difficult to be applied in practical scenarios.We propose to train the model over a number of supervised datasets.Hopefully,experiments show that the improved generalization performance can be achieved for the trained model.