Queen Mary, University of London
We study the problems of visual recognition and tracking of people’s identities and actions in a multiple camera network. Currently, our research focuses on person re-identification(re-id), which matches individual persons across multiple non-overlapping camera views. With re-id, we aim to track a target subject in a large area even though he/she may disappear in some camera views.
The problem of person re-id is very challenging for computer vision, since any person’s appearance would look different due to the variation of view change, lighting change, motion change, resolution change and possibly clothing change. A key problem to be solved is how to learn a model for robust and cross-view appearance-change-tolerant distance matching metric that measures the similarity between person images across views.
In our studies, we identified that the feature discrepancy across camera views is large such that the appearance distributions of different camera views are almost not overlapping. We have solved it by developing a generalised and parametrised feature augmentation for learning view specific feature transforms simultaneously to overcoming view bias and discrepancy on quantifying person appearance features.
Moreover, we investigated that not all people should be tracked but only a few of them on a given watch-list. Hence, we have developed an open-world re-id model that eliminates the impact of non-target people for reducing false verification on target people re-id.
Our current ongoing research explores richer information from depth sensor information in order to cope with severe illumination and clothing change conditions for re-id. We have been exploiting a top-push constraint for reducing the ambiguity of person identity in video motion information. In the coming year, we aim to combine person re-id together with multi-modality action recognition, prediction, and multi-modality fast search, further extending the related works already undertaken last year.