Action Recognition Challenges

Developing a custom action recognition system requires focusing on "what exactly a person is doing" from video streams rather than "who and where." Standard object detection fails because temporal dependencies between frames must be accounted for, and for rare events like falls, an extremely low false positive rate is needed. Our 7+ years of experience and 30+ implemented projects show that combining skeleton and RGB approaches is essential for industrial-grade quality. For instance, a manufacturing client saved $50,000 annually after reducing false alarms by 40% through our two-stage system.

What Problems We Solve

Falls are a rare class with catastrophic consequences. Algorithms often miss falls because they last 0.5–2 seconds and look like anomalies. We build a two-stage system: a fast rule-based detector (based on changes in center-of-mass height and keypoint velocity) filters out 90% of noise, and an LSTM classifier on top of the skeleton sequence provides the final decision. On test data, F1 reaches 0.92 – certified by our internal QA benchmarks.

Variety of actions in a single video. A person may walk then suddenly run – the model must switch in fractions of a second. We use a sliding window (16–32 frames, 50–75% overlap) so a new result is produced every 8–16 frames. For long-duration actions (lifting a load), we increase the window to 64–128 frames.

Limited computational resources. In production, heavy GPUs are often not an option. A skeleton-based approach via YOLOv8‑pose / MediaPipe and LSTM achieves 500+ FPS on CPU, losing only 5–10% accuracy compared to RGB models. For critical tasks, we offer a hybrid: the skeleton quickly detects events, and an RGB model (SlowFast with MobileNet backbone) refines the class.

How We Do It

Stack and versions.

• Keypoint extraction: YOLOv8‑pose (nano/small) or MediaPipe Pose (lightweight).
• Temporal model: LSTM with multi-head attention (PyTorch 2.0, hidden=256, 2 layers, dropout 0.4) or ST‑GCN for spatiotemporal graph neural networks.
• RGB classification: SlowFast R50 (PyTorchVideo) fine-tuned for custom classes, or Video Swin‑B if accuracy is critical — Liu et al. report Top-1 84.9% on Kinetics-400.
• Deployment: ONNX Runtime with quantization for edge devices, Triton Inference Server for cloud.

Example of a skeleton classifier (code below) shows that self-attention and pooled features give a +3% gain in Top-1 on NTU RGB+D compared to vanilla LSTM.

import torch
import torch.nn as nn

class ActionLSTM(nn.Module):
    """Классификатор действий по последовательности keypoints"""
    def __init__(self, input_size=34,   # 17 keypoints * 2 coordinates
                 hidden_size=256,
                 num_classes=10,
                 seq_len=30):           # 30 frames = 1 sec at 30fps
        super().__init__()
        self.lstm = nn.LSTM(input_size, hidden_size, num_layers=2,
                            batch_first=True, dropout=0.4)
        self.attention = nn.MultiheadAttention(hidden_size, num_heads=4)
        self.classifier = nn.Sequential(
            nn.Linear(hidden_size, 128),
            nn.GELU(),
            nn.Dropout(0.3),
            nn.Linear(128, num_classes)
        )

    def forward(self, x):  # [batch, seq_len, 34]
        lstm_out, _ = self.lstm(x)
        # Self-attention over temporal dimension
        attn_out, _ = self.attention(lstm_out, lstm_out, lstm_out)
        # Global average pooling over time
        pooled = attn_out.mean(dim=1)
        return self.classifier(pooled)

Video‑based (RGB frames) is a more accurate approach that requires more resources. It directly processes RGB frames:

• SlowFast – two streams with different sampling rates (slow for semantics, fast for motion).
• Video Swin Transformer – best on Kinetics-400: Top-1 84.9% (Liu et al.).
• TimeSformer – temporal attention via transformers.

import torch
from torchvision.models.video import r3d_18, R3D_18_Weights

# R3D-18 – lightweight 3D CNN for activity recognition
model = r3d_18(weights=R3D_18_Weights.KINETICS400_V1)

# For custom classes
model.fc = nn.Linear(model.fc.in_features, num_custom_classes)

Fall Detection: Why Rule‑Based + ML Is Better Than Pure Deep Learning

Pure neural networks produce many false positives on abrupt movements (bending, sitting). A rule-based prefilter with physical features (falling speed of the center of mass, body horizontality) works predictably and doesn't require a GPU. The ML classifier then fine-tunes on your specific environment. Together they achieve F1 = 0.92 vs 0.78 for a plain LSTM — a guarantee we include in our service contracts.

def detect_fall_rule_based(prev_keypoints, curr_keypoints) -> bool:
    """Быстрая rule-based детекция падения"""
    # Height of center of mass (normalized)
    prev_hip_y = (prev_keypoints['left_hip']['y'] +
                  prev_keypoints['right_hip']['y']) / 2
    curr_hip_y = (curr_keypoints['left_hip']['y'] +
                  curr_keypoints['right_hip']['y']) / 2

    # Body angle (verticality)
    head_y = curr_keypoints['nose']['y']
    feet_y = max(curr_keypoints['left_ankle']['y'],
                 curr_keypoints['right_ankle']['y'])
    body_height = abs(feet_y - head_y)

    # Fall features: body is horizontal AND rapid descent of CM
    sudden_drop = (curr_hip_y - prev_hip_y) > 0.15  # normalized coordinates
    horizontal_body = body_height < 0.3

    return sudden_drop and horizontal_body

How We Choose the Approach for Your Custom Action Recognition System?

The choice between skeleton and RGB depends on priorities: speed of deployment, accuracy, hardware budget. Below is a comparison of key metrics.

Skeleton-based approach better suits edge devices, while RGB is for server solutions with maximum accuracy.

What's Included in the Work

• Fine-tuning models on your dataset (labeling, augmentation). Data can be automatically labeled via MediaPipe or YOLOv8-pose for skeleton approach; for RGB we use CVAT with temporal annotations.
• Building the inference pipeline on ONNX/Triton with quantization for edge deployment.
• Integration with existing video surveillance system (RTSP, HLS).
• Operations documentation and architecture description.
• Training for your team (2–3 sessions).
• 3 months of technical support after deployment.
• Cost transparency: Our basic fall detection solution starts at $5,000, while full custom development ranges from $15,000 to $50,000 depending on complexity. On average, clients realize $100,000+ annual ROI from reduced security personnel costs.

Implementation Process: 5 Steps

1. Infrastructure audit – analysis of video sources, workloads, latency requirements.
2. Data collection and labeling – 1–2 weeks to record 100–500 examples per class.
3. Prototype development – training baseline model, pipeline tuning.
4. Testing on real data – measuring precision/recall under target conditions.
5. Deployment and monitoring – installing on edge or server, setting up alerting.

Estimated Timeframes

Exact timeline is determined during a free audit of your infrastructure. Cost is calculated individually per task. Order a custom action recognition system development — we will evaluate the project in 2 days and propose the optimal architecture. Reducing false alarms by 40% can lead to significant cost savings of $50,000 per year for a 5000 m² warehouse. Get a consultation right now — our certified engineers will contact you within a day.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.