MEFA-MS: Attention-Based U-Net for Pedestrian and Vehicle Detection
Résumé
A Multimodal image Early Fusion with Attention (MEFA) module can be employed for visual object detection, leveraging the robust capabilities of well-established monomodal computer vision models. However, it exhibits poor inference time under real-world usage conditions. The present paper introduces a novel module, referred to as MEFA-MS. This module proposes a U-Net encoder-decoder architecture incorporating spatial and channel attention mechanisms. It aims to reduce inference time while maintaining high precision for pedestrians and vehicles detection. We conduct a comparative study between the proposed MEFA-MS module and the predecessor MEFA module. The proposed approach entails a 2% improvement in mAP50, accompanied by a 50% reduction in inference time for pedestrian and vehicle detection on the DENSE dataset, employing the RT-DETR object detector. Moreover, we demonstrate that the system exhibits superior performance in all weather conditions, achieving a 4% reduction in false positives compared to the previous MEFA module.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |