Computer Vision
For the last two years AnotherTrail’s consulting arm AnotherPeak has successfully crafted and deployed innovative Edge Compute Proof of Concept models for various organisations across Europe.
Whilst certain models offer increased complexity processing, automation and orchestration, more simplistic concepts are emerging like Tiny Edge. My recent Blog looked at Micro Control Units and Digital transformation, but now we turn our attention to Computer Vision, in effect the use of battery driven wireless optical units that process images inset and simply report the outcome.
No cables, no wires, no GDPR.
The basics of Edge/Cloud models
To recap:
- Take a cloud orchestrated edge compute estate and load with ML/AI engines and operation specific policies.
- Link the Edge device to local data sources(sensors, camera’s etc).
- Process and action most of the data locally(based on policies), and report back to the cloud for analytics.
The result is lower latency, reduced bandwidth requirements, sensitive data kept locally for compliance and security benefits and the ability for the edge to act alone in disaster situations.
Artificial Intelligence
Artificial Intelligence(AI) forms part of this Edge compute service, with computer vision offering the most impactful demonstration of the concept.
However, complex visual engines have their challenges
- Crafting an AI vision engine is time/resource consuming
- Requires considerable reference material to train and test the engine
- Is constantly iterating and learning and therefore requires on-going development resource.
An example : Dwell Time : On average Shop Lifters stay stationary at their target longer than regular shoppers.
- Solution : AI engine pushed to edge compute detects human and location pattern and sends alarm directly to in-store staff.
- Problem : Staff restocking goods also dwell creating false positives.
- Iteration : Rework the AI engine to identify shirt colour of target(corporate uniform) using more reference material.
Now, the more complex the AI the more computation resources required.
An example : Road Junction monitoring. In real time utilising multiple existing CCTV cameras, produce data on traffic flows, types of vehicle and pedestrian interaction in all weather and lighting conditions.
Unlike the Dwell Time example such complexity requires high bandwidth data streaming back to cloud compute resources, hence the benefits of Edge are somewhat diluted.
The Chicken and the Egg
You can’t trial computer vision without an engine.
You can’t construct an engine without resources and hence cost.
To bridge this gap AnotherPeak has the ability to produce simple AI vision engines for our PoC client base to illustrate a concept.
However this still costs, so companies need to identify key business drivers before embarking on an AI vision project.
At this point I could leap head long into technology speak and mention cloud orchestration, micro-services, edge compute resources and federated services, but no, let me ask a simple but very important question.
Who would sell Computer Vision?
Too often technology concepts fail to accommodate the available routes-to-market. Be it simple messaging, integration with existing services or pricing.
For Computer Vision the primary route to market is the CCTV/Security companies/Building Services Management. They are well verse in cabling, control room services and increasingly IP based equipment.
Without causing offence they are not Kubernetes gurus, Devops gods or Digitally Twinned masters, and hence new offerings need to be a simple evolution of existing services.
However their world has its boundaries. CCTV cameras require adjacent power sockets, and network cabling and with high resolution imaging some meaty Ethernet switching capacity.
These boundaries have created huge geographical blindspots, and hence opportunity in the form of Tiny Edge
Enter Tiny Edge.
In its basic form Tiny Edge computer vision is :
- A simple, cheap, small, out-of-the-box IP enabled camera.
- It is battery based(3-5 years)
- It takes low/high definition images for situations awareness task.
- It utilises available wireless communications.
- It is trained to do one task , say monitor 7 Segment LED of a refrigeration unit.
- It utilises AI to limit to minimise data transfer(in various forms)
- It communicates with an Access Point which in turn feeds data northbound
Yes, the service provider needs to understand IP, some wireless networking and management, but that is it.
The concept means intelligent vision anywhere(within wireless range) at a price point significantly below current market levels and without the need for cabling/power.
There are no iterative learning processes, no major cloud orchestration, just a simple automate intelligent action that can augment existing traditional CCTV services.
We have our first two Tiny Edge Vision demos in place at AnotherPeak and we’d be happy to discuss in more detail.
Chris @ anothertrail.com