Sensor Processing, How Hard Can It Be?

When I think back back just a few years ago, there were only a handful of devices that noticeably responded to changes in their environment by sensing the presence of something or a change in position. Not that these were the only ones, but these were the ones that were part of my regular “consumer” interaction with the world in my daily life:

• Tilt sensors in game controllers then in tablets and smart phones
• Proximity detection for my smart phone turns the screen off when I have it up to my ear, or turns a light on at home when I enter the garage
• Heart rate monitors were catching on for working out with a wireless connection to a dedicated wristwatch for monitoring
• Automotive airbag deployment

More recently, there’s just been an explosion in these “consumer” devices—spurred in part by low-energy consumption solutions making battery-powered devices actually usable.

• More integrated and comprehensive wearables for monitoring body functions (fitness and medical), typically connected to your smartphone as the monitor
• Thermostats that learn and respond to your presence
• Smart wristwatches (if you think that charging your watch every day is worth the convenience)
• Continuous listening voice activation for hands-free operations
• Face recognition on my smart phone to unlock it
• Collision/proximity detection and driver assistance in cars using imaging, radar, and sonar

With the “IoT” buzzword being applied to pretty much anything with a sensor and some form of communication, companies are producing more and more innovative products that aid productivity, increase safety, or are just plain “cool”.

Thinking about the kind of processing that is going into these devices, there is a huge range of computation required, from environmental sensors at a few samples per second to image processing and feature detection at over 100M samples per second, and everything in-between with motion detection at 100 samples per second and voice activation at over 10K samples per second.

For more details on this concept, view Chris Rowen’s Whiteboard Wednesdays video.

The success of these consumer products depends on the mix of cost, battery life, must-have features, and time to market. All manufacturers focus on one or more of these areas and hope to still be sufficiently differentiated from their competition at product launch.

For any particular product, any or all of the sensors mentioned above could be used depending on the function envisaged by the creator. There’s no single computational element that can be used for all of these and still give the energy and cost efficiency needed for the wide range of possible end products. These devices often need a combination of control and DSP processing but adding separate processors for those functions reduces energy and cost efficiency and can make the product uncompetitive. The alternative may be to design your own, but that is time consuming, costly, and risky. How do you choose something that is flexible enough AND efficient for your product without having to learn and program multiple devices from different vendors?

Take a look at this White Paper on how Cadence is enabling product designers to choose just the level of computation they need from the new scalable Tensilica Fusion DSP for IoT and wearables:

Neil Robinson