Détection des Piétons par Apprentissage Profond de Transformateur

Abstract

Road and highway safety has become a high-priority issue due to the significant increase in road accidents per year. In the scope of our project, our objective is to design an application that detects pedes- trians in road areas. To achieve this, we have combined the power of Transformers and Convolutional Neural Networks (CNNs) to obtain accurate and fast pedestrian detection. We utilized the DETR (Detection Transformer) model because it integrates both CNNs for extracting features from input images and transformers for model training. We re- trained the DETR model specifically for pedestrian detection, resulting in DETR-P, an adaptation of the DETR model focused solely on pedestrian detection. We used the PRW dataset to test and evaluate our system. The performance of our system has been tested on images and videos of varying complexity. The obtained results are very promising but can be improved