Monocular multi-person pose estimation: A survey

Multi-person pose estimation in unconstrained scenarios, with an unknown number of individuals, is a main step towards scene understanding and action recognition. Due to the recent advancements on the architecture of convolutional networks, body part detectors are now accurate and estimate poses in real-time for both single-and multi-person scenes. In contrast, assigning detected body parts to coherent human poses when there are multiple persons interacting is an arduous task. To name a few of the challenges that arise in such scenes: person-to-person occlusion, truncated body parts, and more sources for double counting. Recently, the community contributed towards solving most of them. Hence, it would be interesting to analyze and compile successful approaches from current literature into research trends, and identify possible gaps for future works. To the best of our knowledge, there is no up-to-date review on the main advancements in the field that target this particular set of challenges. This survey fills this gap by reviewing the main breakthroughs on multi-person pose estimation over the last decade and summarizing their impact on the state-of-the-art. Regarding scientific contributions, we propose a novel taxonomy that categorizes the reviewed methods according to their main contributions to the pose estimation pipeline, lists the main datasets and evaluation metrics to train new models, and provides insights on the best entries of publicly available benchmarks. (c) 2021 Elsevier Ltd. All rights reserved.

Automated summary

Multi-person pose estimation faces challenges such as person-to-person occlusion, truncated body parts, and double counting. Recent research has contributed successful methods to address these challenges. There is currently no up-to-date review on the latest advancements in tackling these challenges, and this study fills that gap.