| Abstract |
Multi-pedestrian attribute recognition (MPAR) is an important task for smart city surveillance. Most existing works focus on the scenario with simple background, lacking performance and accuracy for practical deployment. There needs a comprehensive solution for multi-pedestrian attributes recognition which can perform multi-pedestrian detection, pedestrian attribute recognition, pedestrian re-identification in a well-designed inference network to meet both accuracy and performance requirements. This paper proposes a novel end-to-end Dynamic Fitting Multi-Task Network (DFMTN) for multi-Pedestrian attribute recognition (DFMTN4PAR). DFMTN can perform pedestrian detection, pedestrian attribute recognition and pedestrian re-identification in one shot in a single neural network by adopting a multi-task learning strategy, boosted with the proposed dynamic fitting method to prevent gradient explosions. We have extensively evaluated the proposed DFMTN4PAR approach for its accuracy and performance, and compared with the state of art work, which shows that DFMTN4PAR can significantly improve the detection efficiency while ensuring accuracy, and compared with DeepMAR, the DFMTN4PAR approach improves the pedestrian attributes recognition accuracy by 5%-20%, and the detection speed by 30%-60%. The DFMTN4PAR approach has been successfully deployed to 10 smart residential areas in Qingdao, China and runs stably for over a year. © 2023 IEEE. |
