A patch-based affinity measurement method for online multiple pedestrian tracking
Feidie Liang, Sheng Tang, Rui Lv · 2015
The accuracy of affinity measurement is one determining factor for performance of detection-based multiple pedestrian tracking methods. To get high accuracy, most existing affinity measurement techniques require expensive and laborious data annotation for model training when there is no future information in online applications. In this study, we propose to measure affinity based on patches and use discriminability and representability of patches to describe different importance of different patches in the affinity calculation. Both discriminability and representability of patches are online learned in an unsupervised way, so our method is free of data annotation and future information. In addition, motion, size and time information are explored in our method, so entering/leaving pedestrians, short-term occlusions and missing detections can be partially overcome. We evaluate our approach on several public data sets and show significant improvements compared with state-of-the-art methods.