Comparative analysis of object density estimation algorithms for anomaly detection and object counting tasks.

Camera usage in industrial applications is becoming more common in recent times due to its wide-ranging versatility in solving various emerging problems in a production environment. Instances of camera use can be found in diverse tasks, namely quality control, production monitoring, workplace safety, security surveillance, automation and robotics. In addition to this, nowadays, several industrial applications require the use of smart cameras capable of performing built-in computer vision and machine learning operations, thereby expanding their applicability in vision-based solutions. This work proposes a comparison between Machine Learning and Deep Learning models, addressing the problem of manufactured object distribution density for the purpose of adversity detection and object counting. These algorithms can be useful for supervising the correctness of the production procedure and also for obtaining an estimate of the number of objects produced. The solutions explored would range from more traditional computer vision and machine learning-based approaches to custom deep learning models suited for object counting purposes. Furthermore, the proposed study addresses the applicability of such algorithms in a real-time scenario to gain insights into a control problem.