I’ve been in the vision systems business for a long time. Early on, a basic vision system cost $30,000 before any application engineering or configuration. Image resolution was VGA at best—640 x 480 pixels, or 0.3 MPixel. Processing took as long as one second.
Today, smart cameras start at $3,000, or 10% of the price of an early vision system, not accounting for inflation. Image resolution can be as robust as 100 MPixel and above. Processing takes only milliseconds. Camera prices have fallen from thousands of dollars to hundreds of dollars. Lenses still cost about the same, but their imaging capability is way above what was available earlier.
Light sources used to be relatively cheap—incandescent and fluorescent lamps—often from a hardware store. Today, light sources tend to be more expensive and use LEDs, but they have 50,000-hour lifetimes as compared to earlier 1,000-to-5,000-hour lifetimes. Additionally, lights are available in a wide range of shapes, sizes, and colors.
Getting here has not always been smooth sailing. In the early to middle 1980s, there were many startups battling to build sales in a slowly developing market. This led to unfulfilled promises of performance. Many of these nascent vision system companies and integrators did not survive a subsequent market downturn.
Enter Machine Learning and Deep Learning
Now, we see a surge of new machine learning (ML) and deep learning (DL) companies. They showcase some amazing capabilities. The trajectory of these companies, though, gives a feeling of déjà vu: a great number of startups chasing a slowly developing market, leading to promises that likely won’t be met. We can expect a market backlash, the demise of many of these startups, and eventually a technology that matures and becomes a significant part of vision system solutions.
In my opinion, ML and DL are a big thing. But what comes next? We’re seeing the metaverse evolve, along with the rise of IoT, giving, in turn, rise to embedded vision or the Vision of Things (VoT), which will become ubiquitous.
Embedded Vision Changes the Paradigm
Embedded vision is an entirely new paradigm for vision system companies. It will be characterized by very high volume in the tens of thousands to millions of devices. This demands a low cost and low power-consumption design, a different approach to design and production, and greater sensitivity to security and privacy issues.
The pieces of technology are there to sell low-cost camera modules—comprising sensor, lens, and interface—for only a few dollars each and in high quantities. New specialized, low-power consuming processors are appearing on the market every day, achieving ever better price-performance levels. My research suggests that VoT devices will soon be on the market for around $30 and growth rates may be at 20% or more per year due to high volume. Most vision system designers are not acquainted with designing for high-volume, low-cost production and will have to adapt.
Low power is key and, for most vision system designers, a new constraint. The often-stated goal is a peel-and-stick device that will run on its own battery for 10 years. It is debatable when that goal will be achieved. What is clear, though, is that low power means images can’t be streamed over the network. Processing an image takes less power than transmitting it at video speeds. Even though VoT will be connected as part of the metaverse, the devices will not stream video but rather process the images in the device and send data and perhaps an occasional image over the network. Most VoT devices will be event-based sensors.
Privacy and Security Concerns Arise with Embedded Vision
As VoT devices become not only part of our cars but also part of our offices, homes, recreational activities, and even the clothes we wear, security and privacy become absolutely critical. The famed Google Glass was a technical success even though it failed in the market. People who wore them liked them. But the people around someone wearing the Google Glass were very uncomfortable. Designers of VoT devices will have to incorporate empathy into their design process to understand their market and how to make their devices acceptable from a privacy standpoint.
Security is a big issue for every device that is part of the metaverse. So far, security has been a rather small part of the design of vision systems; the volume is so low that it doesn’t attract much attention from the bad actors. As VoT devices, which can give glimpses into our private lives, proliferate, they become more attractive targets for hackers. Designing for security becomes a major part of the development process for VoT devices.
The market for low volume, crafted vision systems will remain strong with decent growth. But embedded vision is the big thing that’s just beginning; it will address volume markets.
Three questions remain. First, will there be a market for VoT devices? All indicators predict a very large, fast growing, and high-volume market. Second, will there be the resources to design VoT devices? Right now, there are very few vision system designers with an appropriate skill set. Companies wishing to address the VoT market will need to learn how to organically grow their development teams. Third, will the VoT market experience the downturns that early vision systems experienced, and that ML is very likely going to experience? It’s hard to tell. A lot depends on how rationally companies approach entering and growing with this new market.
