DiffusionHead:Dense crowd target head detection based on diffusion model

Le résumé fourni par la source

In several domains, including intelligent traffic management systems, population density analysis, security monitoring, and behavior pattern recognition, head detection is essential. However, traditional object detection methods face significant challenges in dense crowd scenarios, including but not limited to occlusion, motion blur, complex backgrounds, small object detection, and image degradation caused by crowd density. These issues often result in unsatisfactory detection performance. To address these challenges, We propose a head detection framework based on diffusion models, called DiffusionHead. It comprises a cascade refinement module, an adaptive conditional generation module, a conditional refinement module, and the proposed Foreground-Enhanced Context-Aware Feature Pyramid Network (FEC-FAFPN). Using feature fusion and contextual information extraction, the model achieves improved detection accuracy and robustness. FEC-FAFPN is the core component of DiffusionHead, responsible for extracting image features. It integrates a feature pyramid network (FPN), a deformable self-calibrated module (DSCM), and a Context-sensitive Prediction Module (CPM). By enhancing foreground objects and incorporating contextual information, it significantly boosts DiffusionHead’s capability to recognize head features, enabling more accurate head detection. Experimental results on the SCUT-HEAD and CroHD datasets demonstrate that DiffusionHead significantly outperforms existing methods. Ablation studies further validate the importance of the Deformable Self-Calibrated Module and the Context-Sensitive Prediction Module in enhancing the performance of DiffusionHead.

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