Table 1 Human tracking algorithms within a camera
Method | Description | Typical techniques | Pros | Cons |
|---|---|---|---|---|
Generative trackers | To estimate each target’s location and correspondence through searching the most similar target candidate with the minimal reconstruction error | KF | Real-time tracking | Subject to linear target state transition and Gaussian noise distributions; apt to lose the tracked target when a target is occluded |
PF | Non-linear/non-Gaussian tracking and multi-modal processing | High computational complexity | ||
KT | Real-time tracking | Cannot deal with long-term total target occlusion | ||
Discriminative trackers | To separate targets from the background through a classifier, and then jointly to establish these targets’ correspondences across frames through a target association algorithm | JPDAF | Multi-target tracking | Subject to data association between a fixed number of tracked targets |
MHT | Variable number of multi-target tracking under occlusion | Vitally high computational requirement | ||
FNF | Variable number of multi-target tracking under occlusion | Cannot effectively deal with long-time target occlusion |