Evaluation of a visual question answering architecture for pedestrian attribute recognition

Abstract

Pedestrian attribute recognition (PAR) ensures public safety and security. By automatically detecting attributes such as clothing color, accessories, and hairstyles, surveillance systems can provide valuable information for criminal investigations, aiding in identifying suspects based on their appearances. Additionally, in crowd management scenarios, PAR enables monitoring of specific groups, such as individuals wearing safety gear at construction sites or identifying potential threats in sensitive areas. Real-time attribute recognition enhances situational awareness and facilitates rapid response during emergencies, thereby contributing to public spaces’ overall safety and security. This work proposes applying the BLIP-2 Visual Question Answering (VQA) framework to address the PAR problem. By employing Large Language Models (LLMs), we have achieved an accuracy rate of 92% in the private set. This combination of VQA and LLMs makes it possible to effectively analyze visual information and answer questions related to pedestrian attributes, improving the accuracy and performance of PAR systems.