Video Pipeline Framework

Aware AI is a design framework created to solve the “Black Box” problem in video surveillance. While most computer vision algorithms are trapped in rigid, non-customizable software, Aware AI transitions the power to build and perform analytics from the researcher to the end-user through a Streamline Processing Framework.

The Core Problem: The Research-User Disconnect

Most vision algorithms fail in the real world due to:

• Data & Scene Variability: Algorithms optimized on static datasets struggle with real-world shifts in lighting and density.
• Limited Models: Users lack the tools to retrain or adapt models for specific tasks without a developer.

Technical Architecture: The Vision Pipeline

The framework treats analytics as a series of non-blocking, unit-processing blocks called Modules. These are connected via a Publish/Subscribe service to form dynamic pipelines.

• Initialization Stage: Fetches streams, loads weights, and prepares the compute environment.
• Processing Stage: Handles the heavy lifting through Source (data generation), Process (inference/logic), and Sink (output/storage) modules.
• Termination Stage: Finalizes operations, such as database uploads or triggering external alerts.

Scalability via Kubernetes & Containers

To handle large-scale camera networks, the framework is implemented as a containerized cluster:

• K8s Orchestration: Each pipeline is deployed as a logical unit in a Kubernetes cluster, sharing networking and storage resources.
• Microservices Approach: Individual modules (e.g., an MCNN crowd density estimator or an SSD object detector) are isolated as containers, ensuring that failed modules are automatically restarted without killing the entire pipeline.
• Distributed Pub/Sub: Allows modules to be deployed across different compute nodes while maintaining real-time communication.

Real-World Use Cases

The Aware AI software features a “Design Analytics” interface where users can drag and drop modules to create custom logic:

1. Crowd Counting Alert: Combines an MCNN density module with a thresholding module to generate riots/protest alerts.
2. No Parking Enforcement: Chains an object detector, a class filter (for cars), and a spatial ROI filter to detect persistent parking violations in restricted zones.