Consistent camera-invariant and noise-tolerant learning for unsupervised person re-identification