Constraining Relative Camera Pose Estimation with Pedestrian Detector-Based Correspondence Filters

Aldea, Emanuel; Pollok, Thomas; Qu, Chengchao · 2019 · Crossref

DOI: 10.1109/avss.2019.8909859

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 accurately estimating the relative pose between synchronized surveillance cameras in urban environments, particularly those with wide baselines and repetitive visual patterns. Standard feature-matching algorithms often fail in these settings due to the homogeneous appearance of pedestrians and the scarcity of distinct static features, leading to unreliable correspondences and large estimation errors. The authors propose integrating high-level semantic constraints—specifically pedestrian detection and re-identification—into a guided matching framework to filter out spurious feature associations and improve calibration robustness. The methodology builds upon a guided matching algorithm that progressively refines the fundamental matrix using video sequences. The authors introduce two filtering strategies to validate feature matches. The first, "box filtering," uses a YOLOv3 pedestrian detector to ensure that matched interest points either both lie within detected pedestrian bounding boxes or both lie outside them. The second, stricter strategy, "re-id filtering," incorporates a person re-identification algorithm (using an OpenCV model) to verify that matches involving pedestrians correspond to the same individual across camera views. This approach leverages object-level consistency to constrain low-level feature matching. Experiments were conducted on two datasets: Dataset 1, a grayscale video of a crowded mosque courtyard with repetitive architectural patterns, and Dataset 2, an RGB video of four volunteers in an open area with few static features. Evaluation metrics included Root Mean Square Error (RMSE) and Maximum geometric error (ME) against manually annotated ground truth. On Dataset 1, box filtering reduced the final RMSE from 0.87 to 0.68 pixels and the ME from 2.73 to 2.23 pixels, representing an 18% improvement in maximum error. On Dataset 2, both box filtering and re-identification filtering outperformed the baseline, with re-identification achieving the lowest ME of 1.88 pixels. The results indicate that while initial estimation errors may be higher due to fewer matches, the filtering strategies stabilize the solution faster and reject approximately 10% of spurious matches. The study concludes that incorporating pedestrian detection and re-identification significantly enhances the accuracy and stability of relative camera pose estimation in challenging urban scenes. By enforcing semantic consistency, the proposed filters effectively mitigate the ambiguity caused by pedestrian homogeneity and repetitive textures. This approach offers a practical solution for calibrating ephemeral surveillance networks with minimal human supervision, improving downstream tasks such as multi-view detection and tracking.

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.

• dms validation