Edge AI in Video Analytics and Surveillance System

Introduction Edge AI

Processing data near the edge of the network, where the data is generated, instead of a centralized data-processing warehouse. Edge computing enables mobile computing and IoT technologies. It makes data and devices more affordable and connected without increasing responsiveness and reducing latency. Let’s take a deep dive into edge computing and understand how the technology works with Akira.ai's solution for business users.

Click here for Overview of Edge AI Implementation

Why do we need Edge AI?

Emerging technologies like deep learning and neural networks, which have revolutionary potential depending on cloud computing, hampers its runtimes increasing massive power requirements.

Forrester Research, an International Informational Technology Firm, reports that “ Latency is becoming an issue as firms try to push more data to software that runs in the cloud or the data center.” As the amount of data increases, it becomes more uneconomical to do all processing centrally.

Edge Computing acts as efficient technology that brings intelligence closer to the place where intelligence is needed and, in return, unleashes the collective power of intelligent devices. As the number of firms grows and increases, central software platforms handling the inflow of data are being pushed to the edge. The main motivations to choose Edge Computing are:

  • Real-time data processing without latency or delay in the transfer of data.
  • Eliminates lag time or allows smart applications to respond to data instant.
  • A large amount of data is processed near sources resulting in reduced internet bandwidth.
  • Eliminates costs ensuring applications to be used in remote locations.
  • Processing data without putting it in the cloud adds security for sensitive data.

Read more about Explainable AI in Manufacturing Industry

Use of Edge AI in Video Surveillance System

Note: Before moving ahead, please visit the following docs. Ebook Till now, you have understood three things:

  • What is Edge AI?
  • Why do we need Edge?
  • And what is a video surveillance system? How is it implemented?
  • This section belongs to running the video surveillance system on Edge.


Popular Edge Devices and their Hardware specifications (For reference to select framework according to Device)

Architecture of Edge AI in Video Surveillance

The overall architecture of running any model includes the following steps:

Considering these steps, here is the solution diagram:

After getting the model ready, here are steps that will be followed:





  • Image Detection Model
  • Data Collection
  • Data Preprocessing
  • Feature Engineering
  • Model development
  • Model Training
  • Model Deployment 
  • Collection of all the Images
  • Labeling of the Images
  • Dividing it into train and test
  • Conversion of Image Data into appropriate format
  • Developing of the algorithm
  • Training of the model
  • Validation of the model
  • Generation of the automated pipeline to deploy the model

To be done on cloud/ machine

  • Image Matching Model
  • Model Development (which can compare different Images)

  • Result and its Validation
  • Defining image descriptor 
  • Indexing Image dataset
  • Defining Image similarity metric
  • Comparison and Searching
  • Developing a pipeline to validate whether we got a match or not from search. 
  • Developing a pipeline for providing person’s information which got a match 

To be done on cloud/ machine

  • Deployment of Edge
  • Deploying model of Edge

  • Model Validation on Edge
  • Generating the results
  • Developing a pipeline to deploy models on Edge.
  • Validating the model on Edge (as we do in QA)
  • Developing a pipeline to generate the results (generation of prediction and generation of results to Image search)

To be done on Edge

  • Running the models of Edge
  • Model Analysis (for maintain operational accuracy)
  • Model Versioning
  • Result Analysis
  • Developing a dashboard on the basis of KPI for model analysis
  • Developing a pipeline by which the model can be versioned (if new model come after training and re-training )
  • Developing a Dashboard with KPIs to do the analysis of the results 

To be done on Edge


Disclaimer: In general, you should:

  • Never use Raspberry Pi for training
  • Deploy the model on Raspberry Pi and Run scoring/ prediction on it.

Usability of the Solution

Related Post