Abstract
Pedestrian detection is a cornerstone of autonomous driving, yet it faces persistent challenges such as background false positives and missed detections in crowded environments. This paper explores the role of contextual information, arguing that while context is beneficial, it can also lead to localization confusion if not properly managed. We propose a pedestrian detection framework that extends proposals to enrich context and employs a supervised attention module to maintain focus on the target instance. Our method achieves an average precision of 74.78% on the KITTI pedestrian benchmark (hard metric level), ranking first among all methods at the time of publication.
The Challenges: Background Confusion and Crowds
Pedestrian detection systems must operate reliably in complex urban scenes. Two primary issues often degrade performance:
- Background False Positives (Hard Negatives): Objects like poles, trees, or traffic signs often share humanoid shapes, leading detectors to misidentify them as pedestrians.
- Crowd Occlusion: In dense crowds, pedestrians overlap significantly. Standard detectors often struggle to produce precise bounding boxes for individuals when they are partially blocked by others.
While researchers often use "context"—the area surrounding a pedestrian—to improve detection, it is a double-edged sword. As illustrated in the figure above, enriching context can help identify a person but may simultaneously cause "localization confusion," where the network becomes distracted by neighboring pedestrians or background noise, making it harder to pin down the exact boundaries of the target.
Proposed Framework: See Extensively, Focus on the Core
The authors propose a two-stage architecture designed to "See Extensively" while "Focusing on the Core."
Stage 1: Multi-scale Proposal
Using a VGG16 backbone, this stage generates initial candidate boxes (proposals) across different scales to handle pedestrians of varying sizes.
Stage 2: Extension and Supervised Attention
This stage contains the core innovations of the paper:
- Proposal Extension: The Region of Interest (RoI) is explicitly expanded to capture more surrounding context, which aids the network in classifying the object.
- Supervised Attention Module: To prevent localization confusion, the authors introduce a supervised mask. During training, the network is forced to learn an attention map where the "core area" (the actual pedestrian) has a high weight (1.0) and the surrounding context has a lower weight (0.5). This "0.5 weight strategy" ensures the network doesn't completely ignore context but prioritizes the core area for precise localization.
Multi-task Loss
The model is trained using a complex loss function that includes classification, bounding box regression, a novel IoU-based regularization term ($L_{reg}$) to optimize coordinates jointly, and a mask loss ($L_{msk}$) to supervise the attention module.
Experimental Results
The researchers evaluated their method on two major benchmarks, demonstrating both high performance and generalizability.
KITTI Pedestrian Benchmark
The framework achieved 74.78% AP on the "hard" metric level, which includes small and heavily occluded pedestrians. This result ranked first among all published methods at the time of the study.
PASCAL VOC2007
To demonstrate generalizability beyond driving scenarios, the detector was tested on the PASCAL VOC2007 person category, achieving 79.16% AP and outperforming general-purpose detectors like Faster R-CNN and SSD.
Ablation Study
Ablation results confirmed the effectiveness of the proposed modules. While the "Extension" alone improved AP by approximately 4%, adding the "Supervised Attention" further boosted performance, particularly in crowded sets where AP rose from 63.13% to 69.06%.
Conclusion
By balancing the need for broad contextual information with a sharp focus on the target individual, this framework addresses the critical "localization confusion" problem in pedestrian detection. While the two-stage process with a ResNet18 sub-network for the second stage is more computationally expensive than single-shot detectors, the significant gains in accuracy—especially in challenging crowded environments—set a new benchmark for the field.
Summary updated