Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition

Source: arXiv:2608.05115v1 [cs.CV], 5 Aug 2026

This paper addresses the underexplored problem of automatic classroom incident recognition from CCTV-style observations, with a focus on privacy preservation and computational efficiency. The authors note that while schools and classrooms are environments intended to support the learning, development, and well-being of children, incidents can occur—including physical altercations, children running and colliding with furniture, and physical abuse by adults. They argue that manual monitoring of continuous video streams requires sustained human attention and is inherently error-prone, creating the need for automated computer vision assistance.

The paper identifies that classroom incident recognition remains underexplored in both academic research and practical deployment, with a lack of datasets, benchmarks, and tailored methodologies for this task. Distinct challenges include "unusual camera viewpoints, since CCTV cameras are often mounted near classroom ceilings; the presence of both adults and young children, who are typically underrepresented in standard action recognition datasets; and domain-specific action categories... which may be absent from conventional benchmarks. These challenges are further compounded by privacy considerations, since classroom monitoring involves children and other identifiable individuals."

The authors list four main contributions:

1. Introduction of a CCTV-view classroom incident recognition benchmark covering safety-relevant interactions and accident scenarios involving pre-school-aged students and adult teachers, containing both synthetic AI-generated videos and real-world recordings.

2. A lightweight, privacy-preserving recognition framework that operates on skeletal pose features rather than raw images or video, reducing reliance on identifiable visual information.

3. Demonstration that the proposed method outperforms existing pose-based action-recognition baselines while requiring fewer computational resources.

4. Demonstration that models trained on our synthetic data generalize effectively to real-world classroom footage.

The paper reviews several research areas:

• Computer Vision in Education: Prior work has focused on engagement detection, attendance tracking, and assessment, but existing work has largely focused on engagement, participation, discourse, assessment, or general classroom observation, leaving the identification of classroom incidents comparatively underexplored.

• Knowledge Distillation: The authors build on the paradigm popularized by Hinton et al., using temperature-scaled soft targets to transfer 'dark knowledge' from a teacher model to a student model.

• Privacy-Preserving Skeleton-Based Action Recognition: Skeleton-based methods are described as a privacy-conscious alternative to RGB-based video analysis, as pose sequences encode human motion while reducing reliance on the appearance, background, and identity cues present in raw video. However, existing privacy-aware skeleton-based methods have primarily been evaluated on general human action benchmarks, leaving their applicability to real-world educational settings comparatively underexplored.

• Video Generative AI: Modern generators (Sora, Kling, Seedance, WAN, LTX) are used for data augmentation and scenario simulation, but generated classroom videos must be used carefully to avoid unrealistic behavior, privacy leakage, and biased incident representations.

Because violent or safety-critical classroom actions are scarce in public data, while collecting such data through staged enactments may pose injury risks to participants, the authors use generative AI to produce realistic synthetic videos depicting the desired action classes in classroom settings. The pipeline involves:

• Action class selection: In consultation with preschool teachers, seven classes were selected: fall, punch, jump, kick, throw, run, sit, plus a background class serving as a meta-class comprising diverse classroom actions that do not fall within the selected incident categories to improve robustness by reducing false positives on non-incident activities.

• Preliminary generator study: Kling, Seedance, Sora, Veo, Wan, and LTX were evaluated for visual realism, temporal consistency, controllability, action fidelity, and their ability to preserve classroom-specific visual properties. Kling and Seedance were selected.

• Uniform assets: A library of Singaporean school uniforms was generated using state-of-the-art generative image models in a human-in-the-loop process, with each asset manually verified against the corresponding real-world uniform.

• Prompt design: The authors "worked with partner preschools and studied sample CCTV footage from their campuses to identify the desired properties of the generated videos, including typical camera placement, viewing angle, covered area, visible classroom objects, and the apparent size of students and teachers."

• Quality inspection: Each generated video was independently reviewed by three authors, with retention requiring unanimous agreement on visual realism, camera-related properties, motion realism, and action semantics. After annotation, all videos and labels were validated by preschool teachers, who served as domain experts. The final synthetic dataset contains 1,296 samples.

Pose data was extracted from anonymized video data from actual preschools, with action labels annotated following the same verification procedure. The final real-world dataset contains 574 samples.

The authors state: all data collection and handling procedures followed the approved institutional ethics protocol, all authors completed the required IRB/human-subjects research training before accessing any human-subject data, real student images were anonymized, and the released dataset will contain only synthetic videos and approved real-world predicted skeleton keypoints, with no personally identifiable or linkable student or teacher information.

The recognition system has access only to human pose trajectories, rather than the underlying RGB video. In deployment, these trajectories may be extracted from classroom footage by the data-holding institution before being provided to the recognition model.

The core insight is that "position trajectories alone may not fully capture the dynamics of an action. Higher-order kinematic quantities, such as joint velocity and acceleration, can encode motion direction, speed, and changes in movement intensity that are not explicit in joint positions. These signals are particularly useful for distinguishing classroom incidents involving rapid or forceful motion, such as punching, kicking, and throwing."

Joint positions are treated as zeroth-order kinematic features. Velocity (first order) is computed as:

v a t = (j a t - j a t-δ) / δ

Acceleration (second order) is:

a a t = (v a t - v a t-δ) / δ

with a general recursive definition for n-th-order quantities.

Pre-processing: Since higher-order kinematic features amplify errors present in lower-order pose estimates, the authors apply temporal preprocessing—missing joint coordinates are filled using linear interpolation over neighboring valid frames and each joint trajectory is temporally smoothed using a Savitzky–Golay filter before computing derivatives.

Each kinematic order is processed by a dedicated backbone (the paper uses STGCN++ architecture with 64 base channels). All backbones share the same architecture but maintain independent parameters, and each is trained on its corresponding input representation. These backbones are pretrained independently, then frozen and fused. The fused multi-order representation is computed via weighted averaging:

ϕ multi-order = (Σ w m ϕ m) / (Σ w m)

with optimal weights w0, w1, w2 = 1.0, 0.8, 0.2.

To enable efficient deployment, the authors distill the multi-order teacher into a substantially smaller zeroth-order student model that operates only on joint positions. This achieves compression along two axes: 1) from a teacher that leverages multi-order kinematic inputs to a student that uses only zeroth-order joint-position inputs; and 2) from a larger teacher network to a substantially smaller student network. The student uses STGCN++ with 32 channels, reducing the parameters by half relative to each teacher branch (and to roughly one-sixth the total teacher size).

The student is trained with a multi-objective loss:

L MTL = λ KD * L KD + λ Cls * L Cls

where L KD is the temperature-scaled KL-divergence distillation loss (with τ = 4, λ KD = 0.3, λ Cls = 1.0) matching the teacher's softened class probabilities, and L Cls is standard cross-entropy with ground-truth labels.

All models were pretrained on NTURGBD large-scale dataset, then trained on 80% of the synthetic dataset and tested on the remaining 20%. Pose extraction used YOLO pose estimator. Accuracy was the primary metric.

Model Accuracy (%)

CTRGCN 65.56

AAGCN 66.39

STGCN++ 68.05

MSG3D 68.88

PoseC3D 70.54

Ours 71.78

The method improves over PoseC3D, the strongest baseline on this dataset, by 1.24 percentage points, and over MSG3D by 2.90 percentage points, while operating at less than one-tenth of the computational cost of PoseC3D. The authors explain the gains as arising from the combination of denoised hierarchical kinematic representations, fused multi-order teacher learning, and efficient teacher-student distillation, which allows the student to benefit from richer multi-order motion reasoning during training while retaining the low computational cost of a single-stream pose-based model during deployment.

Notably, all methods perform far below typical benchmarks—PoseC3D... reaches only 70.54% accuracy on our classroom incident dataset—indicating that classroom incident recognition presents distinct challenges, including viewpoint variation, subject diversity, noisy pose trajectories, and fine-grained differences between safety-related actions.

Models trained only on synthetic data were evaluated directly on the real-world dataset without any fine-tuning. Results:

Model Accuracy (%)

CTRGCN 53.83

PoseC3D 54.36

AAGCN 54.70

STGCN++ 57.32

MSG3D 59.23

Ours 63.41

The method obtains the best real-world accuracy of 63.41%, outperforming MSG3D, the strongest baseline in this setting, by 4.18 percentage points. The authors attribute this to the combination of pose-based abstraction, learned multi-order motion modeling, and multi-objective distillation. They caution that the decrease from 71.78% on synthetic data to 63.41% on real-world data shows that a synthetic-to-real domain gap remains, and results should not be interpreted as showing that synthetic data fully replaces real-world data.

Three ablations were conducted:

1. Multi-order fusion weights: Sweeping weights from 0 to 1 confirmed the optimal setting w0, w1, w2 = 1.0, 0.8, 0.2, showing all kinematic orders provide unique and complementary information.

2. Distillation loss weight: The best performance was obtained at λ KD = 0.3, suggesting that the classification and distillation losses provide complementary supervision.

3. Multi-objective distillation: Hierarchical kinematic features improve over zeroth-order features; distilling a multi-order hierarchical teacher into a single zeroth-order student allows the student to outperform the teacher; and imposing the parameter-reduction condition (Equation 11) further enables the student to outperform the much larger hierarchical teacher. The authors hypothesize that the student internalizes the teacher's multi-order motion reasoning while its smaller size... encourages more effective and transferable representations. Compared with the base STGCN++ model, the proposed method improves accuracy by 4.1 percentage points and macro-F1 by 5.2 percentage points, with gains mainly driven by better recognition of kick, sit, and run, as well as reduced confusion of kick and sit with the other class.

In this pilot study, the authors constructed a CCTV-based classroom incident recognition dataset using generative video AI, complemented by real-world data collection and developed a lightweight, privacy-preserving approach for recognizing classroom incidents. Their approach outperforms existing baselines while requiring fewer computational resources, and demonstrates the strongest zero-shot synthetic-to-real-world generalization. The authors state they will publicly release our dataset, codebase, and supporting tools to the research community to facilitate further research in privacy-aware classroom safety.

Improvements to AI systems:

1. Privacy-preserving incident recognition from skeletal data: Replace RGB-based action recognition with a skeleton-pose-only pipeline. The AI system receives only 2D/3D joint trajectories, eliminating reliance on identifiable faces, clothing, and backgrounds. This enables deployment in schools and other sensitive environments while complying with privacy regulations.

2. Hierarchical kinematic motion encoding: Enrich the input representation by computing zeroth-, first-, and second-order derivatives of joint positions—position, velocity, and acceleration—after temporal preprocessing (linear interpolation for missing joints and Savitzky–Golay smoothing). This improves discrimination of rapid, forceful actions such as punching, kicking, throwing, and falling, which are poorly captured by position-only models.

3. Multi-order teacher ensemble: Train separate graph-convolutional backbones (e.g., STGCN++ with 64 channels) on each kinematic order, then fuse their outputs via learned weights (optimal weights 1.0, 0.8, 0.2). This captures complementary motion cues and improves recognition accuracy without requiring multiple systems at inference time.

4. Multi-objective knowledge distillation for efficiency: Distill the larger multi-order teacher into a compact student model that uses only joint positions and has half the channels (32 vs 64), using a combined loss of temperature-scaled KL divergence (τ=4, λ=0.3) and cross-entropy classification (λ=1.0). The resulting system achieves higher accuracy than the teacher while requiring less than one-tenth the computational cost of strong baselines like PoseC3D.

5. Synthetic-to-real generalization via structured data generation: Use generative video AI to create large-scale, privacy-safe training data with controlled classroom viewpoints, uniforms, action classes, and camera angles, followed by human-in-the-loop quality filtering and domain-expert validation. This enables training incident-recognition models without collecting real violent footage, which is scarce and ethically problematic.

6. Robustness to noisy pose trajectories: Apply temporal interpolation and smoothing before derivative computation, preventing noise amplification. This improves performance on real-world CCTV footage where pose estimation is noisy due to ceiling-mounted cameras, occlusion, and children’s varying body sizes.

7. Reduced false positives via background meta-class: Include a background class containing diverse non-incident classroom actions to suppress false alarms, making the system more reliable for continuous monitoring.

8. Zero-shot deployment capability: Train exclusively on synthetic data and deploy directly on real-world classrooms without fine-tuning, achieving 63.41% accuracy—outperforming baselines by 4+ points. This allows rapid, privacy-safe deployment to new sites before site-specific labeled data is available.

What the improved AI system can do:

• Automatically detect physical altercations, falls, kicks, punches, throws, and unsafe running/collisions in classrooms from CCTV footage, without ever storing or transmitting images of children.

• Operate in real time on edge devices (e.g., Raspberry Pi-class hardware) due to the lightweight student model.

• Distinguish true incidents from normal classroom activities (sitting, jumping, playing) to minimize nuisance alerts.

• Transfer learned safety reasoning from simulated environments to real-world schools with no additional annotation effort.

• Provide a secure deployment pipeline where the data-holding institution extracts poses from video, and only skeletal coordinates are passed to the recognition model—ensuring that no personally identifiable information leaves the premises.

Can computer vision help make classrooms safer? In this pilot study, we investigate privacy-aware and computationally efficient classroom incident recognition from CCTV-style observations. This setting remains underexplored, with limited benchmarks and few methods designed for the privacy, efficiency, and generalization demands of real-world deployment. We introduce a novel hybrid benchmark combining generative CCTV-style videos with real-world classroom pose data, and propose a lightweight, but robust motion-reasoning framework motivated by the observation that many incidents differ more in motion direction, speed, acceleration, and intensity than in pose alone. To that end, our method first constructs hierarchical kinematic representations of human actions. Our method then distills hierarchical, multi-order kinematic reasoning from a large teacher into a much smaller single-order student, enabling efficient per-person inference while preserving expressive motion understanding. Experiments show that our model outperforms substantially larger baselines at less than one-tenth of their computational cost, while also demonstrating stronger out-of-domain motion reasoning and zero-shot synthetic-to-real generalization. We will publicly release the benchmark, codebase, and supporting tools to facilitate further research in privacy-aware classroom safety.