He Zhang

Northwestern Polytechnical University · China

Unclaimed · AI-structured profileORCID OpenAlex

He Zhang (NWPU): edge-native wearable foundation models and video-language datasets for efficient activity recognition

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Output is not scored: the underlying bibliometric record shows h-index and i10-index of 0 with only partial OpenAlex coverage, so this axis — which would measure total publication volume and productivity trajectory — cannot be reliably computed yet.

Impact—

Impact is not scored: citation-based indicators are missing from this snapshot (h-index and i10-index are 0), so the axis — measuring citation uptake, downstream use, and community influence — has insufficient data and should not be estimated.

Focus0/100

The focus score is 0, reflecting an evenly split field distribution: OpenAlex assigns roughly one-third each to computer vision/video-language understanding and two sensor-based HAR categories, while leading topics span multimodal machine learning (7), reinforcement learning in robotics (4), robot manipulation and learning (4), GAN image synthesis (4), and domain adaptation and few-shot learning (4). The portfolio is broad and exploratory rather than concentrated, and may consolidate around wearable HAR plus video-language modeling as the record matures.

Collaboration56/100

With a collaboration score of 56, OpenAlex records 16 distinct coauthors in the recent window, 12 of whom recur frequently: Sizhen Bian, Ruibo Ming, Lei Sun, Deheng Zhang, Jialu Li, Jian Wang, Zhendong Li, Mengshun Hu, Danda Pani Paudel, Luc Van Gool, Jinjin Gu, and Siyu Yuan. The mix spans a wearable-sensing cluster and a computer-vision cluster (Paudel, Van Gool, Gu), pointing to a mid-density, bridge-style pattern that links the sensing and vision communities.

Activity83/100

Activity is the strongest axis at 83: OpenAlex credits 16 works over the recent three-year window, with per-year counts climbing from 4 in 2025 to 12 in 2026, and three 2026 arXiv preprints already reviewed by Evolve. This steep ramp indicates an intensive current phase, though the short cumulative record means sustained multi-year activity is not yet demonstrated.

Reproducibility43/100

Reproducibility averages 43 (reproMean) across Evolve's three reviewed papers. Signals are mixed: Kairos earns credit for strong leakage control, yet EdgeHAR is rated low because no ablation isolates its disentangled sensor-code factorization, and the agentic HAR studio's perfect Shoaib LOSO F1 comes without a non-agent search baseline; Kairos's model-generated annotations also lack human ground-truth validation. Adding ablations, baselines, and human annotation checks would most directly lift this axis.

Research summary

He Zhang is a researcher at Northwestern Polytechnical University (CN) whose work, as indexed by OpenAlex, centers on human activity recognition (HAR) with wearable and edge sensing, extending into video-language understanding. The recent three-year window records 16 works with 16 distinct coauthors, and per-year counts rise from 4 publications in 2025 to 12 in 2026 — an accelerating, early-stage pattern consistent with h-index and i10-index values of 0. Topic clustering assigns the largest share to multimodal machine learning applications (7), with reinforcement learning in robotics, robot manipulation and learning, GAN-based image synthesis, and domain adaptation/few-shot learning at 4 each. Evolve has reviewed three 2026 arXiv papers: EdgeHAR, a 1.41M-parameter edge-native sensor foundation model where He Zhang is first author; an LLM-agent heuristic learning studio for executable, training-free HAR; and Kairos, a 5,420-hour fine-grained video-language dataset. Field distribution splits evenly (~33% each) across computer vision/video-language understanding and two sensor-based HAR categories. Reproducibility averages 43 across the three reviewed papers, with Evolve flagging missing ablations, unvalidated model-generated annotations, and unusually perfect benchmark scores as recurring evidence gaps.

Collaboration pattern

OpenAlex records 16 distinct coauthors in the recent window (collaboration score 56), with 12 recurring partners. The network splits into a wearable-sensing cluster — Sizhen Bian, Ruibo Ming, Lei Sun, Deheng Zhang, Jialu Li, Jian Wang, Zhendong Li, Mengshun Hu — and a computer-vision cluster including Danda Pani Paudel, Luc Van Gool, Jinjin Gu, and Siyu Yuan. First-authorship of EdgeHAR alongside mostly middle positions suggests a hands-on role in the sensing line while contributing to larger vision collaborations.

Research keywords

human activity recognitionwearable sensingsensor foundation modelsedge computingfew-shot transferdisentangled representation learningLLM agentsvideo-language modelinglong-form video understandingtemporal groundingdomain adaptationmultimodal machine learning

Highlights

EdgeHAR: a 1.41M-parameter edge-native sensor foundation model
First-authored 2026 arXiv work introducing a compact (1.41M-parameter) sensor foundation model with disentangled sensor codes for human activity recognition, showing strong few-shot HAR transfer on edge devices; Evolve's review notes the evidence would be strengthened by an ablation isolating the factorization.
Training-free HAR via an LLM-agent heuristic learning studio
Co-authored 2026 arXiv work in which an LLM agent mines heuristic policies for executable human activity recognition, rivaling deep networks while remaining interpretable and edge-deployable; Evolve flags the perfect Shoaib LOSO F1 and the absence of a non-agent search baseline as weak points.
Kairos: 5,420 hours of fine-grained video-language data
Co-authored 2026 dataset paper providing 5,420 hours of long video annotated at 1 FPS over space, time, and dynamics, with strong leakage control; Evolve cautions that the model-generated annotations lack human ground-truth validation.

Timeline

2026First-authors EdgeHAR, an edge-native compact sensor foundation model for human activity recognition (arXiv).
2026Co-authors an agentic heuristic learning studio for executable, training-free human activity recognition (arXiv).
2026Co-authors Kairos, a fine-grained video-language dataset spanning 5,420 hours of long video (arXiv).

Papers in the Evolve catalog

Frequent co-authors

Sizhen Bian×2 Ruibo Ming×1 Lei Sun×1 Deheng Zhang×1 Jialu Li×1 Jian Wang×1 Zhendong Li×1 Mengshun Hu×1 Danda Pani Paudel×1 Luc Van Gool×1 Jinjin Gu×1 Siyu Yuan×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric indicators come from a partial OpenAlex snapshot (identity: openalex) tied to ORCID 0000-0002-9159-5331; the profile is name-resolved from public bibliographic data.
h-index and i10-index are 0 in this snapshot, and the output and impact axes are unscored due to insufficient data — no citation or venue-quality signals are quoted here.
Only three 2026 arXiv preprints reviewed by Evolve are covered in depth; peer-reviewed venues, grants, and awards are outside this scope.
Counts by year cover only 2025–2026, so long-term productivity trends cannot be assessed.