Generalizable Multi-Camera 3D Pedestrian Detection

TL;DR

This paper introduces a multi-camera 3D pedestrian detection method that operates without scene-specific training, utilizing pose-based localization, clique cover fusion, and optional re-identification to achieve state-of-the-art results on the WILDTRACK dataset.

Contribution

The paper presents a novel, training-free 3D pedestrian detection approach that combines pose heuristics, clique cover optimization, and re-identification for improved generalization.

Findings

Achieved a MODA of 0.569 on WILDTRACK

F-score of 0.78 surpassing existing methods

Effective without scene-specific training data

Abstract

We present a multi-camera 3D pedestrian detection method that does not need to train using data from the target scene. We estimate pedestrian location on the ground plane using a novel heuristic based on human body poses and person's bounding boxes from an off-the-shelf monocular detector. We then project these locations onto the world ground plane and fuse them with a new formulation of a clique cover problem. We also propose an optional step for exploiting pedestrian appearance during fusion by using a domain-generalizable person re-identification model. We evaluated the proposed approach on the challenging WILDTRACK dataset. It obtained a MODA of 0.569 and an F-score of 0.78, superior to state-of-the-art generalizable detection techniques.