Computer vision and machine learning approaches on crowd density estimation: A review

Autonomous crowd counting aims to count the number of people in the region and profile their behaviors over time automatically. It is an interest research topic for crowd management, retail store performance analysis, visual surveillance, public space design and intelligent environment. Existing research unto crowd counting yielded many novel publications and has been approached in different angles. These publications produced good results and excelled in single camera view. But on the other hand, viewpoint invariant crowd counting could benefit the surveillance industry by decrease the deployment cost of large cameras network. The goal of viewpoint invariant crowd counting is to learn a mapping from images to count the crowd and then use this mapping in unseen scenes. This paper reviews on the machine learning feature, regression models and the evaluation metric for crowd counting. It covers two main features which are holistic features and local features. This paper also addresses the limitation of current crowd crowding system and discusses the potential techniques for real-time viewpoint invariant crowd counting that possibility to bring the field one step closer towards a “plug and play” system.