Design of Novel Benchmarking System for Power-Efficient Face Detection Algorithm (PE-FDA) in Artificial Intelligence (AI) based Security System

Recently, there have been active studies of video surveillance using artificial intelligence (AI) and security that operates with face detection algorithms (FDA), such as face identity matching system (FIMS) and pedestrian tracking system (PTS). Benchmarking an optimal FDA is one of the important tasks for designing the AI-based security system. However, this AI-based security system suffers from enormous power consumption due to a high frame rate of multiple cameras. For this reason, the AI-based security system needs to find a power efficient face detection algorithm (PE-FDA). To the best of our knowledge, the conventional FDA benchmarking systems (such as using Iou metric) are not optimized with respect to the power efficiency of FDAs. In this paper, we propose a novel benchmarking system for PE-FDA, including power consumption in AI-based security system. We will define the design of benchmarking system and describe its spatial and temporal challenges. (1) In order to solve the spatial challenges, we propose a novel evaluation score, unitized-distance (UD) metric. (2) In order to improve the temporal challenge, we will introduce frame mapping algorithm. (3) our benchmarking system is designed for PE-FDA in AI-based security system. We validated our benchmarking system of PE-FDA using actual video data obtained from a state-of-the-art security system. Thus, this study of our benchmarking systems can allow FDA to be utilized in AI center monitoring system for the future security system.