What is AI Platform?
AI Platform in Cordatus is a comprehensive intelligent video analytics system that enables real-time object detection, classification, and behavior analysis on video streams. From deploying AI models to analyzing detected objects, setting up alarms, and tracking analytics, AI Platform provides enterprise-grade computer vision capabilities with minimal configuration.
It seamlessly integrates pre-trained models and custom AI pipelines, delivering intelligent insights from your camera feeds with the power of NVIDIA DeepStream SDK for high-performance video analytics.
Why AI Platform?
Unlike traditional video analytics solutions, AI Platform combines ease-of-use with enterprise-grade performance:
- GPU-Accelerated Inference: Utilizes NVIDIA DeepStream for high-performance, low-latency video analysis
- Ready-to-Deploy Models: Pre-trained models for common use cases (traffic, people, faces, vehicles)
- Custom Model Support: Bring your own TensorFlow or NVIDIA TAO models
- Advanced Pipelines: Chain multiple models for complex analysis workflows
- Real-Time Analytics: Detect object behaviors with line crossing, ROI filtering, and direction detection
- Intelligent Alarms: Set up model-based and analytic-based notifications
- Multi-Stream Processing: Process multiple camera streams simultaneously within license limits
Who is it for?
Security & Surveillance:
- Security operations centers monitoring restricted areas
- Access control and perimeter security teams
- Loss prevention and forensic analysis teams
Smart City & Traffic Management:
- Traffic monitoring and flow optimization
- License plate recognition for parking and tolling
- Pedestrian and vehicle counting for urban planning
Retail & Customer Analytics:
- Customer demographic analysis (age, gender, emotion)
- Queue management and crowd monitoring
- Shopping behavior analysis and heatmapping
Industrial & Workplace Safety:
- PPE (Personal Protective Equipment) compliance monitoring
- Restricted area access detection
- Safety protocol violation alerts
Core Capabilities
| Feature | Description |
|---|---|
| Pre-trained Models | TrafficCamNet, DashCamNet, PeopleNet, FaceNet, LPDNet, and more |
| Classification Models | Gender, age, emotion detection, vehicle type/make, license plate recognition |
| Custom YOLO Models | Support for YOLO (v5, v6, v7, v8, v9, v10, v11) with automatic TensorRT conversion |
| Custom TF/TAO Models | Support for TensorFlow Object Detection API and NVIDIA TAO Toolkit models |
| AI Pipelines | Chain multiple models for advanced workflows (e.g., demographic analysis, plate recognition) |
| Analytics Rules | Line crossing, ROI filtering, direction detection, entry/exit detection |
| Model-Based Alarms | Set alerts based on detected object counts and classes (frame-based or time-based) |
| Analytic-Based Alarms | Trigger notifications based on object behaviors and analytics rules |
| ROI Preprocessing | Focus inference on specific areas for improved performance |
| Analytics Dashboards | Customizable charts (distribution, time-based, counter, entry-exit) for data visualization |
| Alarm Records | Watch and review alarm events with AI inference overlays and timeline navigation |
What can I do with AI Platform?
Deploy AI Models & Pipelines
- Choose from ready-to-use object detection and classification models
- Upload custom YOLO models (v5, v6, v7, v8, v9, v10, v11) with automatic TensorRT optimization
- Upload custom TensorFlow or NVIDIA TAO models
- Create multi-model pipelines using visual drag-and-drop Model Composer
- Deploy the same model on multiple cameras or create multiple job instances
- Schedule jobs to run at specific times automatically
Process Video Streams
- Run inference on multiple camera streams simultaneously
- Monitor real-time detection results with bounding boxes and labels
- View FPS, bitrate, bandwidth, and resolution metrics
- Process streams from IP, USB, CSI, and GMSL cameras
Analyze Object Behavior
- Add line crossing detection for counting and tracking
- Define ROI (Region of Interest) areas to focus on specific zones
- Detect movement direction and validate against predefined paths
- Track entry and exit events for access control
Set Up Intelligent Alarms
- Create model-based alarms triggered by detection counts (frame-based or time-based)
- Define analytic-based alarms based on object behaviors
- Receive notifications via multiple channels (email, webhook, etc.)
- Apply different alarms to different cameras within the same job
Optimize Performance
- Use ROI preprocessing to run models only on specific areas
- Assign different models to different ROI zones
- Select specific GPUs for inference when multiple GPUs are available
- Monitor device metrics (GPU usage, temperature, memory) during inference
Track and Visualize Results
- View all detected objects with timestamps on the timeline
- Toggle detection result overlays on/off in real-time
- Switch between picture-in-picture and full-screen views
- Access cumulative analytics counts for line crossing, ROI filtering, and direction detection
- Create custom analytics dashboards with distribution, time-based, counter, and entry-exit charts
- Watch and review alarm records with AI inference overlays and slideshow mode
- Use Master Synchronization to view the same event from multiple cameras simultaneously
How does AI Platform work?
- Add Models or Pipelines: Select ready-to-use models, upload custom YOLO/TF/TAO models, or create pipelines with Model Composer
- Start a Job: Deploy model/pipeline to a device with connected cameras
- Configure Settings: Select cameras, GPU, and add optional alarms or schedules
- Monitor Inference: Watch real-time detection results on the stream page
- Add Analytics: Draw virtual lines, ROIs, and direction arrows for behavior analysis
- Set Up Alarms: Configure model-based or analytic-based notifications
- Optimize with ROI: Use preprocessing to focus models on specific areas
- Visualize Data: Create custom dashboards with various chart types
- Review Alarms: Watch alarm records with AI overlays and timeline navigation
- Manage Jobs: Stop, restart, or delete jobs as needed from Web or Client
Key Concepts
- Model: A trained AI algorithm for object detection or classification
- Pipeline: A combination of multiple models working sequentially (e.g., detect vehicle → classify vehicle type → recognize license plate)
- Job: A running instance of a model or pipeline processing one or more camera streams
- Child Job: A historical record of job executions, tracking how many times a job was started and its lifecycle
- Analytics Rules: Virtual elements (lines, polygons, arrows) defining object behavior detection criteria
- Model-Based Alarm: Alert triggered when detected object count meets specified conditions (frame-based or time-based)
- Analytic-Based Alarm: Alert triggered when objects behave according to analytics rules
- ROI (Region of Interest): A defined area where inference is performed, improving performance and focusing on relevant zones
- Bounding Box: Visual rectangle around detected objects (red by default, blue when triggering analytics rules)
- Dashboard Collection: A group of customizable dashboards linked to a specific model or pipeline
- Chart Types: Distribution (pie), Time-Based (bar/line), Counter (analytics-based), Entry-Exit (in/out counting)
- Alarm Records: Historical alarm events with AI inference overlays, screenshots, and video playback
- Model Composer: Visual drag-and-drop interface for creating custom AI pipelines without coding
- YOLO Support: Automatic conversion of YOLO .pt files to TensorRT engines for optimized inference