Evaluation of a facial recognition engines for a surveillance system

Abstract

This paper presents the evaluation of a facial recognition system when considered in an operational context for the South African environment. Often, performance measures, such as recognition rate, are considered outside of context, which does not provide sufficient information for selecting an appropriate recognition engine. Design Science Research (DSR) in conjunction with elaborated Action Design Research (eADR) was used to translate a real-world scenario to an abstracted research problem. The goal of DSR is to provide an artefact, which in this case is presented in the form of a workable realworld system architecture with a cost-effective recognition engine. Three experiments were conducted on commercial recognition engines, as well as on an open-source engine to determine the performance on MUCT and a subset of the LFW data sets, as well as on a self-constructed database of localized data. Finally, costs were compared between open-source and commercial recognition engines, and a scenario was put forward to show the cost-effectiveness of these two systems in a specific example. Ultimately, it is known that the performance of a facial recognition system is affected by the performance of the recognition engine, but it was evident that the total system must be considered before an informed decision on a recognition engine can be made