[Paper Review] Reciprocal Distance Transform Maps for Crowd Counting and People Localization in Dense Crowd.

Dingkang Liang, Wei Xu+2 more|arXiv (Cornell University)|Feb 16, 2021

Video Surveillance and Tracking MethodsComputer Science4 citations

TL;DR

This paper proposes Reciprocal Distance Transform (R-DT) maps to improve crowd counting and people localization by replacing traditional density maps with a novel representation that eliminates overlap and preserves precise head locations. The method achieves state-of-the-art performance in localization and competitive results in counting, with strong generalization across datasets.

ABSTRACT

In this paper, we propose a novel map for dense crowd counting and people localization. Most crowd counting methods utilize convolution neural networks (CNN) to regress a density map, achieving significant progress recently. However, these regression-based methods are often unable to provide a precise location for each people, attributed to two crucial reasons: 1) the density map consists of a series of blurry Gaussian blobs, 2) severe overlaps exist in the dense region of the density map. To tackle this issue, we propose a novel Reciprocal Distance Transform (R-DT) map for crowd counting. Compared with the density maps, the R-DT maps accurately describe the people location, without overlap between nearby heads in dense regions. We simultaneously implement crowd counting and people localization with a simple network by replacing density maps with R-DT maps. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art localization-based methods in crowd counting and people localization tasks, achieving very competitive performance compared with the regression-based methods in counting tasks. In addition, the proposed method achieves a good generalization performance under cross dataset validation, which further verifies the effectiveness of the R-DT map. The code and models are available at this https URL.

Motivation & Objective

• To address the limitation of regression-based crowd counting methods in precisely localizing individuals in dense crowds.
• To overcome the inherent blur and overlap in density maps caused by Gaussian smoothing and high density.
• To develop a new map representation that maintains accurate spatial localization of individuals even in highly overlapping regions.
• To enable simultaneous, accurate crowd counting and person localization using a unified deep learning framework.

Proposed method

• The R-DT map is constructed by applying a reciprocal distance transform to the ground-truth head annotations, which assigns each pixel a value based on the inverse of the distance to the nearest head.
• Unlike conventional density maps, the R-DT map produces sharp, non-overlapping peaks at head locations, eliminating the blurring effect of Gaussian kernels.
• The method replaces standard density map supervision with R-DT map supervision in a CNN-based network for joint counting and localization.
• The network is trained to regress the R-DT map, enabling precise localization through peak detection and counting via peak counting or integration.
• The approach leverages the geometric properties of the reciprocal distance transform to preserve spatial fidelity in dense regions.
• The framework is trained and evaluated using standard crowd counting datasets, with cross-dataset generalization tested.

Experimental results

Research questions

• RQ1 Can a novel map representation improve the accuracy of person localization in dense crowd scenes compared to standard density maps?
• RQ2 Does the use of Reciprocal Distance Transform maps reduce localization ambiguity caused by overlapping heads in dense regions?
• RQ3 How does the proposed method perform in comparison to state-of-the-art regression-based and localization-based crowd counting methods?
• RQ4 To what extent does the R-DT map generalize across different datasets without fine-tuning?
• RQ5 Can a single network simultaneously achieve high performance in both crowd counting and person localization using the R-DT map?

Key findings

• The proposed R-DT map achieves state-of-the-art performance in person localization on standard crowd counting benchmarks.
• The method achieves competitive counting accuracy, matching or exceeding that of leading regression-based methods.
• The R-DT map significantly reduces localization errors in dense regions due to the absence of overlapping Gaussian blobs.
• The model generalizes well under cross-dataset validation, demonstrating robustness and transferability.
• Peak detection on the R-DT map enables precise person localization, with localization accuracy outperforming existing localization-based methods.
• The method maintains high performance even in extremely dense scenarios where traditional density maps fail due to blurring and overlap.