Abstract
Person Search aims at retrieving a specific individual (the query) within a collection of whole scene images from diverse, non-overlapping cameras. It has the potential to play a pivotal role in various public safety applications like suspect searching and identifying abandoned luggage owners. Person Search encompasses two Computer Vision challenges: 1. Object Detection, which entails localizing humans in whole scene images, and 2. Person Re-Identification, where the query image is compared with images of detected individuals to establish identification. The critical nature of Person Search underscores the imperative to safeguard it against security threats, such as adversarial attacks, which can result in non-detection or misidentification. While adversarial attacks and defense mechanisms have been extensively studied for both Object Detection and Person Re-Identification, there is a noticeable gap in research concerning Person Search. This work presents a comprehensive Systematic Mapping Study and taxonomy of adversarial attacks and defenses in Person Search, utilizing Parsifal and ChatGPT 4 for in-depth analysis. We highlight the persistent challenges associated with Person Search and discuss prospects for future advancements in addressing its vulnerabilities.