Passenger Face Characteristic Classification in the Urban Rail Transit Based on YOLO-ResNet50 Cascading Model

Le résumé fourni par la source

To accurately track dynamic passengers in the rail transit environment and precisely identify passenger facial features including gender recognition and age estimation, this paper proposes the YOLO-ResNet50 cascading model. The algorithm first uses YOLO-V5 for dynamic passenger face tracking; secondly, it introduces the constant acceleration Kalman filter to improve the DeepSORT algorithm for passenger face numbering; finally, it uses the ResNet50 model to recognize the facial features of numbered passengers. The difference from existing face recognition algorithms is that the YOLO-ResNet50 cascading model proposed in this paper can achieve unique recognition of numbered faces, overcoming the multiple repeated recognitions caused by passenger movement. This study can not only reduce computational power consumption but also improve the accuracy of facial feature recognition. Experiments show that the proposed algorithm achieves a Mean Absolute Error (MAE) of 2.16 years for age estimation and a gender recognition accuracy of 99.45% on the AFAD (Asian Face Age Dataset). In actual rail transit scenarios, the age estimation MAE is 3.05 years and the gender recognition accuracy is$\mathbf{9 4. 8 1 \%}$for frontal unobstructed faces; the miss detection rate (MDR) in subway carriages is 17.83 %. The results demonstrate that the algorithm has good robustness and practicality. The accurate identification of passenger facial features can provide differentiated services for rail transit and support commercial decision-making.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.