Ai-MicroCloud™ for IoT

Video Surveillance-as-a-Service

Video Analytics-as-a-Service

 

Overview:

Driven by plummeting camera prices and the recent successes of computer vision-based video inference applications, organizations are deploying cameras at scale.

The following trends will provide opportunities to video surveillance market participants:

  • Increased government focus on public safety

  • Need to utilize and examine unstructured video surveillance data in near real time

  • Potential significant crime reduction due to surveillance cameras

  • Growing need among enterprises to leverage Business Intelligence (BI) and actionable insights for advanced operations

  • Limitations of manual video analysis

  • Increased use of artificial intelligence and deep learning for video surveillance systems

 

Video analytics enhances video surveillance systems by performing near real-time event detection tasks, post-event analysis, and extraction of statistical data, while reducing manpower costs and increasing surveillance system effectiveness.

Intrusion systems, integrated with video analytics, offer more accurate automated surveillance. If an intruder triggers a pre-defined security rule, the automated intrusion management system generates an alert, thereby enabling timely and precise responses.

 

Municipalities must increase surveillance of public places, such as parks, landmarks, bridges, monuments, critical infrastructure to minimize threats to citizens, support law enforcement activities and enhance smart cities management.

Value of Video data.JPG

Solution Brief

Problem:

Scaling video analytics to massive camera deployments, presents new and mounting challenges, as compute cost grows proportionally to the number of camera feeds. Processing video feeds from large deployments, requires a proportional investment in high-performance compute (HPC) hardware (GPUs) or cloud resources (GPU machine time), the costs of which can easily exceed those of the camera hardware.

 

Inference accuracy and cost of processing are the two key metrics in video analytics applications. Inference accuracy is a function of the analytics model used and the labeled data, used for training and video characteristics, such as frame resolution and frame rate. These factors also influence the cost of processing – larger models and higher quality videos enable higher accuracy, while requiring the additional cost of increased resource consumption or accepting higher processing latency. When video feeds are analyzed at an Edge or cloud cluster, the cost also includes the bandwidth cost of sending the videos over a network, which increases with the number of video feeds.

 

AI workload management becomes challenging when video streams from multiple cameras are used by different algorithms to draw inferences simultaneously, e.g., a vehicle tracking algorithm detects a car coming into a gas station and a human action predictive algorithm identifies a person getting out of the vehicle and carrying a suspicious object, like a gun. Pre-processing of video data and Orchestration of compute resources for AI inferences is key to delivering a highly accurate and cost-effective video surveillance solution.

AI Micro Cloud with Video Analytics Algorithms

The Solution:

Zeblok Computational provides a robust technology platform that enables development and implementation of AI-powered video analytics solutions. The solution accommodates video feeds in a variety of formats & resolutions, ranging from a few cameras to thousands of cameras. Zeblok’s Artificial Intelligence/Machine Learning Platform-as-a-Service (AI PaaS) enables data scientists and other AI developers to develop their models using unique curated algorithms from Zeblok’s Intelligence Marketplace and integrate them with enterprise applications easily and seamlessly. Zeblok deploys its AI PaaS to data centers, public clouds and Edge locations, essentially creating a private AI micro cloud for each client.

Service Providers can offer Video-Surveillance-as-a-Service to their clients by developing bespoke AI models and provide sophisticated AI-based Video-Analytics-as-a-Service to a variety of clients responsible for monitoring facilities, such as schools, highways and public spaces to identify vandalism and other suspicious behavior, monitor and analyze traffic patterns, identify traffic violations and many more.

AI Video Analytics diagram v1.3.jpg
Architecture

There are 3 main parts in the architecture of this solution.

  1. Video streams from cameras at edge locations

  2. Zeblok AI-PaaS, with relevant AI algorithms in its Intelligence Marketplace

  3. Integration with enterprise applications

Video streams from Cameras at the Edge

Video cameras installed at Edge locations stream data 24/7. AI models need historic datasets for training purposes. Once the model is trained, inferences can be processed at each edge location or at the Core data center. Edge locations must be equipped with appropriate compute, storage, and networking capabilities to handle the volume of data coming in from multiple cameras and to provide inferences at the edge or at the central command center depending on the client applications.

Zeblok AI PaaS

Zeblok Computational’s AI PaaS provides all the composable foundational components that data scientists and other AI developers need, with a simple user interface and familiar frameworks that enables them to get started on model development in minutes, leverage familiar open-source frameworks and original curated AI algorithms, that are ready to be integrated into the model. Zeblok’s powerful orchestration engine enables seamless scalability of HPC and an accelerated data lake. Zeblok AI PaaS can be deployed to any data center on any hardware. Our AI runtime environment provides APIs that can be easily integrated with client applications.

Our proprietary orchestration layer software supports multi-cloud, multi-class scheduling, allowing simple multi-user AI/ML workstations to co-exist with HPC workloads required for substantial AI/ML model training. It permits on-demand usage of GPU capacities for AI/ML model development. It provides a rapid prototyping environment to promote AI models as APIs for rapid consumption into enterprise business processes.

Zeblok, through its affiliation with National Science Foundation’s CVDI, has access to unique AI algorithms in the video analytics space that are available for model development, providing functionality such as:

  • Early event detection & predictions

  • Human action recognition

  • Visual counting

  • Hand detection

  • Suspicious object detection

  • Facial behavior analysis

  • Suspicious behavior detection

  • Deep learning for illumination and material analysis

Enterprise Applications

The foundational utilities on the PaaS layer of the Zeblok AI PaaS acts as bridge between the video analytics platform and enterprise applications, such as:

  • Intrusion Detection: Perimeter protection, boundary monitoring, restricted zone monitoring

  • Suspicious Activity: Loitering, fighting, vandalism

  • Traffic & Parking Management: Illegal parking, one-way violation, speeding, no helmet, traffic congestion, vehicle counting

  • Smoke & Fire Detection: Early warning of smoke and fire in open areas and indoors

  • Crowd Counting: Crowd counting and flow analysis

  • Vehicle Identification: License plate detection and recognition

  • Non-Invasive Body Temperature Monitoring: Passive thermal imaging for temperature monitoring

  • Dwell Time: Amount of time customer spends in specific areas

  • Social Distancing and Usage of PPP: Workplace safety & hygiene compliance, managing occupancy, identifying protocol violators

Benefits
Camera Compatibility

Connect to an existing infrastructure including IP cameras, Analog cameras, NVR’s and DVR’s

Video Pre-Processing

Perform preprocessing including data sampling, image, and clip extraction from any connected camera to make AI scalable and cost efficient

Scalable & Optimized for Video Analytics

An integrated AI PaaS that can scale from few cameras to thousands of cameras located across multiple geographical locations to build AI models to suit any need - near real-time, periodic, or on-demand video analytics

Zeblok AI-WorkStation

Data scientists can begin AI/ML model development in minutes using AI-WorkStation that provides out-of-the-box CUDA optimized AI frameworks with popular data science language bindings, such as R, Scala, and Python.

Intelligence Marketplace for Algorithms

Growing library of proven original AI algorithms available within AI-WorkStation. Curated via closed loop validation ecosystem, enabling customized AI model development to suit any requirement.

Turnkey High Performance Computing Orchestration

Deploy AI workloads to single or hundreds of Nvidia GPUs with a single click making it easy to work on large volume of datasets for AI model development

Cloud Native architecture

Turnkey cloud native AI PaaS provides instant usability and seamless scalability, with flexibility to enable additional services

 Multi-Cloud deployment

Flexibility to deploy composable foundational components of Zeblok AI PaaS within the MSSP’s data center or within third party cloud service providers like AWS, GCP, Azure or IBM

Agile Edge deployment

Hardware agnostic autonomous provisioning of edge infrastructure with flexibility to push AI inferences to any Edge location provides agility to provision single or thousands of Edge locations globally