The Challenge of a Production Ready Hardware for TinyML

Hi there, I’m posting this to share my experience when trying to implement a TinyML project working in a small/middle sized company and as well to hear any advice from the community.

We are working in a computer vision project, I’m not allowed to share all the details, but basically we need to keep track of vehicles inside a car workshop.

We implemented a simple visual wake word application + apriltag detection using the OpenMV platform, and it works very well with the development board and wifi shield.

The problem we are facing now is how to migrate this to a production ready hardware. There are many factors to pay attention to when implementing this, for instance, the hardware will be outdoor (at least most of it), so an IP65+ is needed.

First we tried to make custom hardware based on the OpenMV designs, but then we realized that the volumes required for it to be financially viable are just to high. 3D printing 300 outdoor cases is too expensive, injection molds are cheaper but you need to be in hundreds of thousands units volume-wise. Plus the chip shortage makes it very difficult for a small sized company to make custom hardware.

I found some production ready hardware using a Jetson Nano and other linux platforms, but those are very overkill for this application.

Do you know any outdoor camera/device that can be programmed with TinyML on it?

I think at this point (sadly) we will do this with the classic cloud approach.

You say: “using a Jetson Nano and other linux platforms, but those are very overkill for this application. I think at this point (sadly) we will do this with the classic cloud approach.”

Have you done a cost estimation between the production-ready Linux platform implementation, the COTS you have found on the market and the Cloud-based implementation approach? Your hardware can be overkill but maybe, in the long term, your deployment cost (cost Cloud, connectivity, etc) will be lower.

1 Like

That’s right! The cost will be equal after about 5 years, and from that point on, the cloud cost will be higher. That was an argument I made, however is harder to sell a project that requires a high initial investment.

I understand your situation, best engineering solution vs. business decision (cost of investment).

I don’t know your specific requirements, as you said you can not share the details. But, if you don’t find a programmable camera, maybe a solution is to buy low cost COTS Wifi Cameras, stream the video data to a Linux platform, for example, Advantech has some edge AI solution which is based on Jetson Nano [MIC-710AI AI Inference System based on NVIDIA® Jetson Nano™]. (MIC-710AI - AI Inference System based on NVIDIA® Jetson Nano™ - Advantech) and perform the inference on the edgeAI system. Maybe Advantech has some other more suitable solutions for your use case?

Regards,
J.

1 Like

Thanks for the advice @JrV, I think that could be a viable alternative.