Crowd Counting in Still Images for RoboSoldier: A Survey

Abstract

Part of the ultimate key capabilities for a RoboSoldier will lie in its ability to count the total number of people within a crowd with pin-drop precision. Crowd Counting (CC) relates to the estimation of the existing number of objects within a still image or video frame. Crowd counting has been successfully used in a wide range of applications from gathering business intelligence, such as consumer shopping patterns and ensuring normal operating conditions, to urban planning for crowd safety and stability through traffic monitoring. However, crowd counting has a fair range of its challenges like, high cluttering, varying illumination, severe occlusion, Perspective Distortion (PD), Irregular Object Distribution (IOD), non-object scales, and Location Information Loss (LIL). The emergence of Fast Learning Algorithms (FLA) for Deep Belief Networks (DBN) and Deep Learning (DL) techniques triggered groundbreaking research interests due to their inherent capabilities of overcoming shortfalls of traditional algorithms which were based on hand-designed feature extraction techniques. However, crowd counting is now sliced into two clear and distinct techniques namely traditional and Convolutional Neural Network (CNN). Traditional techniques can be broadly classified into those that count by regression and those that count by using density estimation techniques. The more recent and advanced is CNN whose architecture has data training capabilities drilling down to the base of as many network layers as available. Categorisation of these layers has been used to extensively offer researchers sound edifice in designing new and powerful distributed control algorithms as well as monitoring algorithms applicable to a wide range of crowd counting applications concerning public gatherings, military defense, wildlife census, etc.