Multi-View Pedestrian Recognition:
This work was performed at the University of California: Santa Cruz under Professor Hai Tao and in co-operation with Doug Gray. Parts of this work were published at PETS 2007. The published paper can be found here.
The purpose of this work was to develop techniques and metrics to solve the pedestrian re-acquisition problem using appearance information only. There has been significant work on pedestrian re-acquisition for tracking using spatio-temporal information, but relatively little work in re-acquisition using appearance information.
The difficulty with pedestrian re-acquisition in multi-camera systems is that the orientation of the pedestrian is likely to be different in each camera view. In other words, you may see the front of the pedestrian in one view, and the back in another. Furthermore, their clothing may not be identical in both views due to things like backpacks, purses, and complicated clothing patterns.
Our method attempts to build a same/not-same classifier using AdaBoost with pairs of haar-like box filters (one from each image) as the weak classifier. To this end, we have created a database of several hundred pedestrians seen from two or more views. This dataset can be found here.
We publisehd aspects of our work regarding the dataset and standard performance metrics for such appearance-based multi-view recognition systems at PETS 2007, along with a few details of our method at recognition. The paper can be found here.