Gesture Recognition
Gesture recognition identifies intentional hand, arm, body, face, or object motions such as swipes, waves, signs, nods, pointing, or interaction commands. It is a specialized form of temporal action recognition, often with tighter latency, smaller motion regions, and stronger user-interface constraints than generic action recognition.
The input representation determines the method. A gesture system may use RGB clips, cropped hands, pose-estimation keypoints, skeleton tracks, depth, inertial sensors, optical-flow, or fused streams. The same product may combine a fast keypoint baseline for obvious gestures with a deep video model for ambiguous or occluded cases.
Method families
Gesture recognition methods fall into several families:
| family | typical input | mechanism | useful when |
|---|---|---|---|
| Template matching / dynamic time warping | Keypoint or sensor trajectories | Compare a trajectory with stored examples under time warping or distance rules. | Gestures are simple, few, and latency must be very low. |
| Handcrafted features + classical classifier | Keypoints, optical flow, shape descriptors | Compute velocities, angles, histograms, or flow summaries and classify with SVMs, random forests, HMMs, or logistic models. | Data is limited and interpretability/debugging matter. |
| Skeleton sequence models | Body or hand keypoints over time | Feed joint coordinates, bone vectors, or graph edges into RNNs, temporal CNNs, or graph convolutional networks. | Pose is reliable and background appearance should be ignored. |
| RGB/depth deep video models | Cropped clips or full frames | Learn spatiotemporal features with 3D convolutional networks, two-stream models, or depth/RGB fusion. | Appearance and motion jointly define the gesture. |
| Transformer-based video models | Patch, tubelet, keypoint, or track tokens | Use video transformers, masked-video pretraining, or foundation-model features over hands, objects, and temporal context. | Long-range context, transfer learning, multi-person interaction, or language-facing outputs matter. |
| Spatio-temporal action detection | Frames, actor boxes, tubes, or track tokens | Predict actor/action tubes and class labels jointly instead of requiring a fixed tracker-then-classifier split. | Multi-person videos need both “who moved?” and “what gesture happened?” |
| Online trigger models | Streaming frames, keypoints, or logits | Estimate whether enough evidence has arrived to fire a command. | The system must respond before the gesture fully completes. |
For tracked keypoints , a simple trajectory baseline is mean velocity:
A template classifier can compare and richer trajectory features against gesture prototypes. A deep model instead learns a representation from raw frames, keypoints, or fused streams and predicts
Real systems also need person tracking and track aggregation so one user’s gesture is not mixed with another person’s motion.
Pose-Guided Person RoI Pipeline
In multi-person gesture videos, the first problem is usually not the gesture class; it is assigning evidence to the right person. A common applied pipeline is:
- Detect people with a detector or pose model.
- Track each person over time with a tracker.
- Extract fixed-length person RoI clips around each track.
- Run a gesture recognizer on each person RoI clip.
- Attach the prediction back to the person track and timestamp.
This pipeline changes the input domain seen by the recognizer:
| input domain | benefit | risk |
|---|---|---|
| Full frame | Keeps body, object, and scene context. | Small hands can be diluted by background and other people. |
| Person RoI track | Removes other people while preserving body, hand, and nearby object context. | The hand can still be small inside the crop. |
| Static hand crop | Preserves a stable coordinate frame around a hand region. | Crop choice can miss body context or the object being manipulated. |
| Tracking hand crop | Keeps the hand large and centered over time. | Motion cues can be removed because the coordinate frame follows the hand. |
| Sparse token keep | Saves computation by keeping person- or hand-region video tokens. | Geometry can break if kept tokens lose their original positions. |
Hand RoIs are useful when hand localization is reliable and the gesture is mostly local to the hand. Person RoIs are the safer default for multi-person videos because they keep the gesture tied to the tracked actor and preserve body context. For video foundation models used as frozen backbones, these input-domain choices can dominate the probe result. If the checkpoint is frozen, a lightweight head measures whether the representation already contains the needed gesture information; it does not teach the backbone a new gesture concept.
Modern Method Families
The mean-velocity template baseline above is useful for intuition, but it is not representative of modern gesture-recognition systems. Current systems usually separate the question “what signal should represent the gesture?” from the question “does the model also need to localize the actor?“
| method family | output | where it fits |
|---|---|---|
| Trajectory templates and dynamic time warping | A label for one tracked keypoint or sensor sequence. | Low-data command gestures with simple motion geometry. |
| Skeleton GCNs and keypoint transformers | A label from joint coordinates, bone vectors, or hand landmarks. | Pose is reliable and background appearance should be suppressed. |
| 3D ConvNets, two-stream CNNs, and SlowFast-style video CNNs | A clip or RoI label from RGB, depth, optical flow, or fused streams. | Local spatiotemporal filters are a good inductive bias, especially when motion and appearance jointly define the gesture. |
| Video transformers, masked-video models, and video foundation models | A clip, RoI, token, or probe-head label from patch or tubelet tokens. | Long context, transfer from large pretraining, or input-domain comparisons are important. |
| Spatio-temporal action detectors and tubelet models | Actor tubes plus action or gesture labels. | Multi-person videos need actor localization, temporal extent, and classification to be coupled. |
The last family matters for person-tracking workflows. A practical system can use a detector, tracker, and per-person classifier, but that is not the only formulation. End-to-end spatio-temporal action-detection models predict action tubes or actor-token trajectories with class labels, so localization and recognition are trained together. This can reduce hand-built aggregation rules, but it usually needs annotated boxes, tubes, or actor-level supervision and is harder to debug than a modular tracker-plus-classifier pipeline.
Model Selection
| requirement | likely starting point |
|---|---|
| Few command gestures on a fixed camera | Keypoint rules, template matching, or lightweight classical classifier. |
| Hands are small but pose is reliable | Hand detector plus keypoint sequence model. |
| Multi-person scene | Person detection, tracking, per-person RoI clips, then gesture classification; or an end-to-end spatio-temporal action detector when actor/action tube supervision is available. |
| Gesture depends on object appearance | RGB/depth model, two-stream model, or 3D convolutional network. |
| Gesture unfolds over many frames | Temporal CNN, recurrent model, transformer, or sliding-window recognizer. |
| Multiple people may gesture at once | Tracking plus per-track gesture classification, or an action-tube model that jointly localizes actors and labels gestures. |
| Low-latency UI command | Streaming model plus trigger-point prediction. |
Caveats
Gesture labels are culturally, linguistically, and product specific. Hands are frequently occluded, small, motion-blurred, or outside the crop. Pose models can fail on unusual viewpoints, while RGB models can overfit to background, clothing, or camera placement. End-to-end action-tube models can remove brittle tracker/classifier handoffs, but they need richer actor-level labels and more careful evaluation of identity switches, duplicate tubes, and missed short gestures. A trigger-point prediction system must avoid firing before the gesture is distinguishable from ordinary movement.
References
- Kapitanov et al., 2022, HaGRID: HAnd Gesture Recognition Image Dataset
- Cao et al., 2018, OpenPose: Realtime Multi-Person 2D Pose Estimation
- Tran et al., 2015, Learning Spatiotemporal Features with 3D Convolutional Networks
- Bertasius et al., 2021, Is Space-Time Attention All You Need for Video Understanding?
- Liu et al., 2021, Video Swin Transformer
- Wang et al., 2023, VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking
- Wang et al., 2024, InternVideo2: Scaling Foundation Models for Multimodal Video Understanding
- Yan et al., 2018, Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition
- Zhao et al., 2022, Multi-Scale Spatial Temporal Graph Convolutional Network for Skeleton-Based Action Recognition
- Zhao et al., 2021, TubeR: Tubelet Transformer for Video Action Detection
Nav