A Study for Attention-Based Multi-View Pedestrian Detection in Crowded Environments

다국어 초록 (Multilingual Abstract)

Detecting pedestrians in crowded scenes remains a challenging problem due to severe occlusions, overlapping appearances, and significant viewpoint variations. Although multi-camera systems provide complementary observations to alleviate these issues, most existing multi-view approaches rely on explicit camera calibration and bird’s-eye-view （BEV） projections. Such geometry-dependent designs are sensitive to calibration inaccuracies, prone to feature distortion, and often exhibit limited generalization across diverse camera configurations and environments. In this study, we present a robust calibration-free multi-view pedestrian detection system that performs joint reasoning in image feature space without the need for camera settings or ground-plane transformations. Instead of relying on predefined geometric mappings, our method learns multi-view feature relationships through a fully learnable transformation mechanism that embeds view-specific representations into a shared latent space for effective multi-view interaction. To further enhance representation quality, we introduce a Squeeze-and-Excitation augmented feature fusion module for adaptive multi-view reweighting, along with a CBAM-enhanced refinement module that improves both spatial and channel-wise attention for better localization under high occlusion. Additionally, a zero-initialization gating strategy is incorporated to stabilize feature integration and enable progressive multi-view learning. Extensive experiments on Wildtrack, MultiviewX, and MVPerception demonstrate that our approach achieves state-of-the-art Multi-Object Detection Accuracy （MODA） scores of 95.5%, 96.6%, and 94.5%, respectively, while maintaining competitive localization performance. These results validate the effectiveness and generalization capability of calibration-free multi-view feature learning in complex real-world environments.

목차 (Table of Contents)

• CHAPTER 1 1
• INTRODUCTION 1
• 1.1 Overview 1
• 1.2 Problem Statement 3
• 1.3 Contributions 4
• 1.4 Organization of Thesis 5
• CHAPTER 2 7
• LITERATURE REVIEW 7
• 2.1 Overview 7
• 2.2 Calibrated BEV-based multi-view pedestrian detection 8
• 2.3 Calibration-free multi-view detection 10
• CHAPTER 3 12
• PROPOSED METHOD 12
• 3.1 Overview 12
• 3.2 Feature Extraction Network 13
• 3.3 Feature Projection Matrix Estimation 14
• 3.4 Squeeze & Excitation Augmented Feature Fusion 15
• 3.5 Fusion Back-Projection to View-Specific Feature Spaces 17
• 3.6 CBAM-Enhanced Feature Refinement 17
• 3.7 Zero-Initialization Gating Strategy 18
• CHAPTER 4 19
• EXPERIMENTAL RESULTS AND DISCUSSION 19
• 4.1 Overview 19
• 4.2 Experimental Setup 19
• 4.3 Evaluation metrics 19
• 4.4 Datasets 21
• 4.4.1 Wildtrack Datasets 21
• 4.4.2 MultiviewX Dataset 22
• 4.4.3 MVPerception Dataset 23
• 4.5 Quantitative Result 24
• 4.6 Qualitative Result 27
• 4.7 Ablation Study 29
• 4.7.1 Effect of SE Channel Attention 31
• 4.7.2 Effect of CBAM-Based Feature Refinement 32
• CHAPTER 5 33
• CONCLUSIONS AND FUTURE WORK 33
• 5.1 Overview 33
• 5.2 Limitations and Future Work 34
• BIBLIOGRAPHY 35