LMGait: Language-Guided and Motion-Aware Gait Representation for Generalizable Recognition

Zhengxian Wu1,†, Chuanrui Zhang1,†, Shenao Jiang1,†, Hangrui Xu2, Zirui Liao1, Luyuan Zhang1, Huaqiu Li1, Peng Jiao1, Haoqian Wang,✉️1
1Tsinghua University     2Hefei University of Technology
†Equal contribution     ✉️Corresponding author
{zx-wu24,zhang-cr,jiangsa24}@mails.tsinghua.edu.cn, wanghaoqian@tsinghua.edu.cn
Paper
Code
Models
Citation

Abstract

Gait recognition enables remote human identification, but existing methods often use complex architectures to pool image features into sequence-level representations. Such designs can overfit to static noise (e.g., clothing) and miss dynamic motion regions (e.g., arms and legs), making recognition brittle under intra-class variations.

We present LMGait, a Language-guided and Motion-aware framework that introduces natural language descriptions as explicit semantic priors for gait recognition. We leverage designed gait-related language cues to highlight key motion patterns, propose a Motion Awareness Module (MAM) to refine language features for better cross-modal alignment, and introduce a Motion Temporal Capture Module (MTCM) to enhance discriminative gait representations and motion tracking.

Proposed Method

Key Ideas

  • We leverage language cues as an explicit signal to guide the vision encoder to focus on key regions of motion. This approach enables our method to extract more representative gait features. As far as we know, this is the first attempt to introduce language-guided learning into the gait recognition task.
  • To better align textual and visual features in the latent space, we design a Motion Awareness Module (MAM) that dynamically adjusts the textual features.
  • We design a Motion Temporal Capture Module (MTCM) that captures the walking process of pedestrians from two complementary perspectives: the pixel-level and the region-level.
Proposed method overview
Pipeline of the proposed LMGait, it consists of five components. Specifically, the video input is processed through the frozen Dinov2 model for feature extraction. The text query guides the network to focus on gait-relevant regions, and it is aligned with the image feature space through the frozen CLIP text encoder and the fine-tuned MAM module. The Representation Extractor generates diverse features, while the Motion Temporal Capture Module captures posture changes during walking. Finally, the extracted features are input into the Gait Network for recognition.

🎥 Multimodal Gait Representation with Visual–Language Priors

We introduce a multimodal gait recognition pipeline that jointly leverages visual observations and language-based semantic priors. By injecting domain-specific motion descriptions into visual feature learning, the model is guided to attend to gait-discriminative body regions, improving robustness under cluttered backgrounds and occlusions.

🧠 Motion-Aware Language Modulation

Instead of treating language features as static prompts, we propose a Motion Awareness Module (MAM) that adaptively refines textual representations based on gait dynamics. This enables the language branch to emphasize motion-relevant semantics while suppressing distractive cues, softly modulating visual features without introducing rigid constraints.

⏱️ Language-Guided Temporal Motion Modeling

To capture the continuous nature of human walking, we design a Motion Temporal Capture Module that jointly models pixel-level and region-level motion patterns. Benefiting from language-guided visual representations, the temporal module aggregates motion trajectories more effectively, avoiding noise accumulation and enabling stable, discriminative gait modeling over time.

Module Analysis

Module analysis figure (1) Module analysis figure (2)
Module analysis. Fixed language prompts and direct temporal aggregation on raw visual features often fail to adapt to changing motion cues, leading to noise accumulation and weakened discrimination. By introducing motion-aware language modulation and language-guided temporal capture, our framework dynamically emphasizes motion-relevant semantics and aggregates pixel-level and region-level motion trajectories more effectively, resulting in more stable and discriminative gait representations.

Results

Representation Method Testing Datasets
CCPG SUSTech1K
CL UP DN BG Mean NM CL UF NT Mean
skeleton GaitGraph2 (Teepe et al. 2021) 5.0 5.3 5.8 6.2 5.6 22.2 6.8 19.2 16.4 18.6
Gait-TR (Zhang et al. 2023a) 15.7 18.3 18.5 17.5 17.5 33.3 21.0 34.6 23.5 30.8
GPGait (Fu et al. 2023) 54.8 65.6 71.6 65.4 64.2 44.0 24.3 47.0 31.8 41.4
SkeletonGait (Fan et al. 2023b) 40.4 48.5 53.0 61.7 50.9 55.0 24.7 52.0 43.9 50.1
Silhouette GaitSet (Chao et al. 2022) 60.2 65.2 65.1 68.5 64.8 69.1 61.0 23.0 65.0 18.6
GaitBase (Fan et al. 2024) 71.6 75.0 76.8 78.6 75.5 81.5 49.6 76.7 25.9 76.1
DeepGaitV2 (Fan et al. 2023a) 78.6 84.8 80.7 89.2 83.3 86.5 49.2 81.9 28.0 80.9
SkeletonGait++ (Fan et al. 2023b) 79.1 83.9 81.7 89.9 83.7 85.1 46.6 82.5 47.5 81.3
Sil+Parsing+Flow MultiGait++ (Jin et al. 2024) 83.9 89.0 86.0 91.5 87.6 92.0 50.4 89.1 45.1 87.4
RGB GaitEdge (Liang et al. 2022) 66.9 74.0 70.6 77.1 72.7 - - - - -
BigGait (Ye et al. 2024) 82.6 85.9 87.1 93.1 87.2 96.1 73.3 93.2 85.3 96.2
LMGait(ours) 84.8 87.0 88.5 93.6 88.5 96.4 79.8 93.9 87.0 97.1
Comparison of state-of-the-art methods on the CCPG (Li et al. 2023) and SUSTech1K (Shen et al. 2022) datasets

Citation

@misc{wu2026languageguidedmotionawaregaitrepresentation,
      title={Language-Guided and Motion-Aware Gait Representation for Generalizable Recognition}, 
      author={Zhengxian Wu and Chuanrui Zhang and Shenao Jiang and Hangrui Xu and Zirui Liao and Luyuan Zhang and Huaqiu Li and Peng Jiao and Haoqian Wang},
      year={2026},
      eprint={2601.11931},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.11931}, 
}
×
This website is inspired by Absolute Zero. We sincerely thank the authors for their contributions to the community.