Human Behaviour
Analysis
AI system that understands human behavior from video — activity recognition, facial biometrics, sentiment detection, personal space interaction analysis, and engagement tracking. Built for safety, healthcare, research, and retail.
What It Does
Multi-layered behaviour understanding from video input to actionable intelligence.
Activity Recognition
Identifies complex human activities using temporal sequence modeling. SlowFast Networks and Video Swin Transformers process video at dual temporal scales to recognize actions like walking, sitting, reaching, and multi-step task sequences.
Facial Biometrics
FaceNet-powered facial embedding extraction for identity verification, age/gender estimation, and expression classification. Operates in real-time with privacy-preserving embedding storage — no raw images retained.
Sentiment Detection
Multi-modal sentiment analysis combining facial expression recognition, body language cues, and vocal tone (when audio is available). Classifies emotional state across a continuous valence-arousal spectrum.
PSI Detection
Personal Space Interaction analysis tracks proximity between individuals, detecting engagement zones, conversational clusters, and spatial violations. Uses calibrated depth estimation and zone regression.
Interaction Analysis
Maps interaction patterns between people — who talks to whom, engagement duration, group formations, and social dynamics. Tracks zone markers, proximity events, and engagement scoring over time.
Data Fusion Engine
Aggregates outputs from all sub-systems — detection, tracking, activity, sentiment, PSI — into unified behavioral profiles. Feeds real-time alerts, prediction models, and NLP/LLM summarization layers.
System Architecture
The full processing pipeline from video input to behavioural intelligence.
| Stage | Technology | Details |
|---|---|---|
| Video Input | RTSP / Camera SDK | Live video stream acquisition from multi-camera setups, frame extraction and normalization |
| Person Detection | YOLOv8 / YOLOv10 | Body detection, bounding box extraction, keypoint mapping, identity assignment |
| Task Detection | ATSS / Conditional DETR | Fine-grained activity detection, object-interaction recognition, compositional reasoning |
| Temporal Modeling | SlowFast / Video Swin | Dual-rate temporal analysis — slow pathway for spatial, fast pathway for motion dynamics |
| Activity Recognition | Compositional AI | Multi-step activity classification, event detection, anomaly identification |
| Facial Biometrics | FaceNet / PyTorch | 128-dim face embeddings, expression classification, age/gender estimation |
| Sentiment Analysis | Multi-Modal Fusion | Expression + body language + context for valence-arousal sentiment mapping |
| PSI Detection | Zone Regression | Personal space interaction detection, proximity measurement, engagement zone mapping |
| Interaction Analysis | Graph Networks | Social graph construction, conversation tracking, group dynamics, engagement scoring |
| Data Fusion | Custom Engine | Unified behavioral profile aggregation from all sub-system outputs |
| NLP / LLM Layer | GPT-4o / Claude | Event summarization, natural-language reports, anomaly explanation, alert generation |
| Blockchain Verification | Immutable Ledger | Data integrity verification, audit trail, tamper-proof event logging |
| Production Pipeline | Kafka + Docker | Real-time alerts, predictive analytics, monitoring dashboards, horizontal scaling |
How Behaviour Is Analyzed
Actions Detected
Body pose, movement patterns, and activities identified from video
Identity Mapped
Facial embeddings and biometric features extracted per person
Sentiment Scored
Expression, body language, and context fused into sentiment
Interactions Tracked
Social dynamics, proximity, and engagement patterns analyzed
Reports Generated
LLM summarizes events, flags anomalies, generates alerts
Use Cases
Safety & Security
Monitor facilities for unusual behavior, falls, altercations, or unauthorized access. Real-time alerts for security teams with LLM-generated incident summaries.
Retail Analytics
Understand customer movement, dwell times, engagement with displays, and staff-customer interactions. Optimize store layouts and staffing decisions.
Healthcare & Elder Care
Track patient mobility, detect falls, monitor engagement levels, and alert caregivers to behavioral changes. Privacy-preserving design with no raw image storage.
Research & Training
Quantify human behavior for academic research, ergonomic studies, and workforce training programs. Structured data export for statistical analysis.
Tech Stack
Want behaviour intelligence for your space?
We deploy behaviour analysis systems for healthcare, retail, security, and research environments. Let's talk about your use case.