Course Objective & Scope
Knowing that a person is in the frame is very different from understanding what they are doing: fall detection, assembly-step verification, unsafe movement or violence recognition all require temporal context. This advanced programme compares 3D CNN architectures (SlowFast, X3D), video transformers (TimeSformer, VideoMAE v2) and pose/skeleton graph networks (MMPose + ST-GCN).
Short-clip classification is separated from temporal action localisation in untrimmed videos, with hands-on training in PyTorchVideo and MMAction2. Participants build an action-recognition prototype and data-collection protocol for their own field scenario, such as EHS compliance or standard operating procedure checks.
Who Should Attend?
- AI engineers developing behaviour, motion and anomaly analysis on video
- Data scientists working on workplace safety (EHS), manufacturing ergonomics and sports analytics
- R&D teams building deep-learning camera solutions
- Prerequisite: PyTorch and experience with CNN/Transformer models in computer vision
Key Takeaways
- Choose appropriately between RGB-based (SlowFast, VideoMAE) and skeleton-based (ST-GCN) approaches
- Feed 2D/3D human pose estimation outputs into action classification models
- Locate action start and end times in untrimmed streams via sliding windows and temporal localisation
- Apply augmentation and class-imbalance strategies for rare anomalies and fall events
- Optimise clip length, stride and compute budget for real-time field inference
Curriculum
Temporal Modelling and Action Recognition Tasks
- Clip classification, spatio-temporal action detection and temporal localisation
- Structuring Kinetics, AVA and custom enterprise video datasets
3D CNNs and Two-Stream Architectures
- Spatio-temporal convolution in I3D, SlowFast and X3D
- The cost of optical flow and RGB-only alternatives
Video Transformer Models
- Divided space-time attention with TimeSformer and VideoMAE v2
- Fine-tuning self-supervised video backbones on enterprise data
Pose Estimation and Skeleton-Based Action Recognition
- 2D/3D keypoint extraction with YOLO-Pose and MMPose
- Background-invariant motion analysis and ergonomic scoring with ST-GCN
Temporal Localisation and Anomaly Detection on Live Streams
- Catching events in untrimmed streams with ActionFormer and sliding windows
- Latency, false-positive control and edge-device constraints in the field
Other Courses in This Category
-
Shared embedding spaces with CLIP/SigLIP, Qwen2.5-VL, LLaVA and InternVL architectures, visual QA, multimodal RAG and QLoRA fine-tuning, for engineers building systems that reason across images and text.
- Live Course
- 5 modules
-
Surface defect detection on production lines, unsupervised anomaly detection with anomalib (PatchCore, PaDiM, EfficientAd), industrial camera/lighting design and PLC integration, for quality and manufacturing engineering teams.
- Live Course
- 5 modules
-
Inpatient fall prevention, ICU monitoring, hand-hygiene and PPE compliance, operating-room instrument counting and privacy-preserving edge anonymisation, for hospital digital transformation and healthtech teams.
- Live Course
- 5 modules
-
DICOM/NIfTI processing, 2D/3D lesion and organ segmentation on CT/MRI/X-ray with MONAI and nnU-Net, whole-slide pathology (WSI) analysis and clinical validation, for medical imaging and biomedical AI teams.
- Live Course
- 5 modules
-
Reception and lobby queue/crowd analytics, buffet tray replenishment and food-waste measurement, car-park ANPR, pool/facility safety and privacy-compliant anonymous analytics, for hospitality technology teams.
- Live Course
- 5 modules
-
Model conversion with ONNX Runtime, TensorRT, OpenVINO and TFLite, INT8 quantization, camera pipelines on NVIDIA Jetson and Raspberry Pi, GStreamer and OTA updates, for teams running vision models at low latency in the field.
- Live Course
- 5 modules
Get in touch
