IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences
Online ISSN : 1745-1337
Print ISSN : 0916-8508
Extracting Re-Identification features from image space for end-to-end multi-view pedestrian tracking framework
Taiga YAMANE, Satoshi SUZUKI, Tomohiro TANAKA, Mana IHORI, Naoki MAKISHIMA, Shota ORIHASHI, Ryo MASUMURA
Author information
JOURNAL FREE ACCESS Advance online publication

Article ID: 2025EAP1201

Details
Abstract

Multi-view pedestrian tracking (MVPT) aims to track pedestrians in the form of a bird's eye view (BEV). In MVPT, re-identification-based (ReID-based) end-to-end methods are one of the most effective approaches. They detect pedestrians and extract ReID features representing their appearances within a single model. These methods generate a BEV feature by projecting image features from multiple views into the BEV space via a perspective transformation, and then perform pedestrian detection and ReID feature extraction using that BEV feature. One problem with these methods is that the pedestrian body structure collapses in the BEV feature due to distortion caused by the perspective transformation. This prevents the model from fully capturing appearance information on pedestrians, leading to insufficient ReID features. To overcome this problem, we propose a new ReID-based end-to-end MVPT method, named MRIS. MRIS captures pedestrian appearance information from each view image space, where the pedestrian body structure is preserved. To achieve this in end-to-end MVPT, MRIS leverages back-projection from the BEV space into each view image space, which is the inverse projection of the perspective transformation. First, MRIS back-projects the BEV feature and detected pedestrian locations in the BEV space to each view image space. Back-projecting pedestrian locations enables the model to extract appearance features, which represent the appearance information, from image features for each pedestrian and each view. Back-projecting the BEV feature complements the information of occluded parts in each view image, helping the model extract effective appearance features. Then, since pedestrian appearance varies from one view to another, MRIS aggregates appearance features from multiple views into a single ReID feature for each pedestrian. Extensive experiments demonstrate that MRIS outperforms previous ReID-based end-to-end methods and show that capturing appearance information from the image space leads to better tracking performance than doing so from the BEV space.

Content from these authors
© 2026 The Institute of Electronics, Information and Communication Engineers
Previous article Next article
feedback
Top