Table 4 Human tracking algorithms across non-overlapping cameras

Human re-id

To identify whether a human taken from one camera is the same as one taken from another camera or not

Feature extraction

Extracting discriminative and robust visual features help to improve human re-id accuracy

Difficult to find suitable feature combination to effectively describe human appearance

Distance metric learning

Learning a distance metric helps to mitigate cross-view human appearances’ variations.

Require manually pairwise labeling of training data

CLM-based tracking

To track humans through establishing the link (correlation) models between two adjacent or among multiple neighboring cameras

Supervised learning-based CLM

Easy to establish and learn CLM

Unfeasible to scale up to large-scale camera networks due to a mass of manually labeled efforts

Unsupervised learning-based CLM

Help to achieve self-organized and scalable large-scale camera networks due to no need of human labeling efforts

Estimated CLM may decrease the accuracy due to higher outlier percentage.

GM-based tracking

To track humans through partite graph matching based on input observations (detections, tracklets, trajectories, or pairs)

MAP optimization solution framework

Human tracking in complex scenes such as occlusion, crowd, and interference of appearance similarity

It is difficult to get the optimal solution.