View in EDS

Bias-driven prediction update network for long-term 3D human motion prediction.

Saved in:
Bibliographic Details
Title: Bias-driven prediction update network for long-term 3D human motion prediction.
Authors: Zhong, Jianqi1 (AUTHOR) janki@sziit.edu.cn, Tang, Junyuan2 (AUTHOR) 2023280197@email.szu.edu.cn, Yang, Yixin2 (AUTHOR) 2210433005@email.szu.edu.cn, Tan, Xu1 (AUTHOR) tanxu1981@163.com, Xie, Tianming1 (AUTHOR) tmxie@sziit.edu.cn, Cao, Wenming2 (AUTHOR) wmcao@szu.edu.cn
Source: Multimedia Systems. Aug2026, Vol. 32 Issue 4, p1-18. 18p.
Subjects: Movement sequences, Forecasting, Convolutional neural networks, Encoding
Abstract: 3D human motion prediction is essential for human–machine systems such as human–-computer interaction, autonomous driving, and intelligent robotics, where long-term prediction for future motion is particularly vital for safety and seamless interaction. Existing approaches, particularly those based on autoregressive architectures, often suffer from cumulative error propagation, where small inaccuracies in early frames amplify over time, leading to severe degradation in long-term prediction. Moreover, current sequence modeling methods frequently lose fine-grained temporal information, further limiting prediction continuity and robustness. To overcome these issues, we propose a Bias-Driven Prediction Update Network (BD-PUNet), a novel non-autoregressive framework designed to suppress error accumulation while preserving temporal detail. BD-PUNet incorporates a Prediction Update Module that progressively refines motion sequences, together with a Bias-aware Feedback Mechanism that feeds deviation-based bias features back into the network. This design stabilizes early frames and mitigates downstream bias propagation. A complementary position-based branch-Positional Encoding-based Temporal Convolutional Network explicitly captures fine-grained positional cues, which are fused with Discrete Cosine Transform-based encodings to yield robust and complementary temporal representations. Extensive experiments on the benchmark datasets Human3.6M, CMU-Mocap, and 3DPW demonstrate the state-of-the-art performance of our method. [ABSTRACT FROM AUTHOR]
Copyright of Multimedia Systems is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:3D human motion prediction is essential for human–machine systems such as human–-computer interaction, autonomous driving, and intelligent robotics, where long-term prediction for future motion is particularly vital for safety and seamless interaction. Existing approaches, particularly those based on autoregressive architectures, often suffer from cumulative error propagation, where small inaccuracies in early frames amplify over time, leading to severe degradation in long-term prediction. Moreover, current sequence modeling methods frequently lose fine-grained temporal information, further limiting prediction continuity and robustness. To overcome these issues, we propose a Bias-Driven Prediction Update Network (BD-PUNet), a novel non-autoregressive framework designed to suppress error accumulation while preserving temporal detail. BD-PUNet incorporates a Prediction Update Module that progressively refines motion sequences, together with a Bias-aware Feedback Mechanism that feeds deviation-based bias features back into the network. This design stabilizes early frames and mitigates downstream bias propagation. A complementary position-based branch-Positional Encoding-based Temporal Convolutional Network explicitly captures fine-grained positional cues, which are fused with Discrete Cosine Transform-based encodings to yield robust and complementary temporal representations. Extensive experiments on the benchmark datasets Human3.6M, CMU-Mocap, and 3DPW demonstrate the state-of-the-art performance of our method. [ABSTRACT FROM AUTHOR]
ISSN:09424962
DOI:10.1007/s00530-026-02366-y