An evidential framework for pedestrian detection in high-density crowds

Vandoni, Jennifer; Aldea, Emanuel; Le Hegarat-Mascle, Sylvie · 2017 · Crossref

DOI: 10.1109/avss.2017.8078498

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of pedestrian detection in high-density crowd scenes, characterized by heavy occlusion, visual homogeneity, and the absence of distinct background features. Standard detection methods often fail in these contexts due to the small size of targets and the inability of individual detectors to capture sufficient discriminative information. The authors aim to identify the most effective visual descriptors for crowded environments and propose an evidential fusion framework that outperforms existing statistical learning fusion methods by explicitly modeling spatial imprecision. The proposed method utilizes four distinct descriptors: Histogram of Oriented Gradients (HOG) for shape, Local Binary Patterns (LBP) for texture, Gabor filter banks for frequency and orientation, and Daisy for stereo-matching-like orientation maps. Each descriptor feeds into an independent Support Vector Machine (SVM) classifier. To fuse these probabilistic outputs, the authors employ belief function theory, specifically the Transferable Belief Model. This framework allows for the representation of both uncertainty (derived from classifier scores) and spatial imprecision (derived from the spatial heterogeneity of uncertainty values within a neighborhood). The fusion process uses a conjunctive combination rule on basic belief assignments, which are defined using morphological operators to account for the spatial context of each pixel. Experiments were conducted on high-density crowd images acquired during the Hajj pilgrimage in Makkah. The results demonstrate that the four descriptors exhibit strong complementarity: HOG and Gabor provide high recall but lower precision due to false positives, while LBP and Daisy offer higher precision but lower recall. The evidential fusion method effectively combines these strengths, achieving superior performance compared to Multiple Kernel Learning (MKL) and naive probability product fusion. Precision-recall curves indicate that the proposed approach yields higher overall precision and recall than the baseline methods. Visual analysis confirms that the fusion reduces false positives and correctly identifies heads even in cluttered regions, although detection of dark veils remains a persistent challenge. The significance of this work lies in its demonstration that modeling spatial imprecision through belief functions provides a robust solution for fusing detectors in difficult surveillance contexts. By leveraging the complementary nature of diverse visual features and handling uncertainty more effectively than standard statistical approaches, the framework offers a reliable method for pedestrian detection in high-density crowds, with potential applications in security and crowd monitoring systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• signal detection theory