How Advanced AI Algorithms Can Unlock the Power of PIR Sensor Data


David J Klein, PhD

Chief Scientist, 2predict, Inc.

Motion sensors have been around since the 1950s when Samuel Bango applied the fundamentals of radar to ultrasonic waves to create a basic detector for burglar alarms. Since then, they’ve been used in a wide variety of applications, from microwaves to touchless entry to surveillance cameras.

But the technology hasn’t changed much until recently. Today, sensors that feature advanced AI algorithms leveraging machine learning (ML) are being developed, to enable higher detection accuracy and broader functionality. This article takes a look at the mechanics and typical uses of PIR sensors, and how AI can expand their capabilities.

What Is a PIR Sensor?

A passive infrared sensor (PIR sensor) is an electronic sensor that measures infrared (IR) light radiating from objects in its field of view. They’re commonly used to power motion detectors in things like household appliances or gadgets — for example, a home video surveillance camera.

PIRs are made of a pyroelectric sensor that can detect levels of infrared radiation. In other words, they measure temperature changes in different regions of a given area that heat-emitting objects cause. Everything — a person, an animal, even a package — emits low levels of radiation, and the warmer something is, the more radiation is emitted, changing the surrounding temperature.

PIR sensors are designed to detect that temperature change. One part of the sensor “warms up” creating a change pulse that’s detected by the other half of the sensor and triggers the programmed response. In a typical deployment, a device may contain multiple PIR sensors. Lenses are used to focus light from the area being monitored onto the sensors.

However, most of the time, PIR sensors have thresholds set to simply detect any movement and trigger a response — not to give information about the object that moved. Your security camera will turn on when someone — or something — approaches your home, but it can’t tell what’s approaching. Is it a human? An animal? A vehicle making a three-point turn in your driveway? Designed to work in low-power applications, they’re too simple to distinguish between moving objects; all they can do is turn on and off in response to motion.

Since PIR sensors can’t tell if the moving object is human — or more specifically, an unwanted visitor — they produce high rates of false positives. In addition, lowering the threshold for the triggers results in missed detections.

Using DNNs to Enable Actionable Detections

Simple on/off algorithms can detect motion but ignore other information such as complex patterns of sensor change over time, or how the sensor channels relate to one another. Recently, advanced AI algorithms for ultra-low-power computing environments have been developed to address this issue, unlocking more information from PIR sensor data, as well as reducing false positives and missed detections.

Currently AI algorithms can use ML models such as deep neural networks (DNNs) to mine information collected by PIR sensors and provide additional detections. They can even distinguish different types of moving objects based on the sensor output patterns across multiple sensor channels.

Here’s how: DNNs leverage much longer time spans of sensor data from multiple channels simultaneously and can learn how this complex sensor data differentiates the presence of different kinds of objects (such as humans, animals or cars). They can also be trained to determine the speed of an object’s motion, and its size and position. To be able to learn these patterns, ML algorithms are trained using many examples of sensor data traces in response to different kinds of events that are important for the given application.

A properly trained DNN-based detector can greatly reduce false alarms and missed detections in a human vs. vehicle detection task. In one test, data engineers trained a DNN-based detector using 200 1-minute PIR time series from a wide variety of locations. When compared to the results of a simple threshold and count algorithm, it reduced false alarms by 80%, missed detections by 60%, and both types of errors simultaneously by 40%.

Although DNN-based PIR detection algorithms have until recently been too compute-intensive for low-power environments, that’s changing. New solutions feature lower power consumption — in some cases as low as 200 microwatts during active sensor processing.

Beware of Latency

Because DNNs record longer time spans of sensor outputs, signal latency can be an issue. This is especially true in video applications, where video data is buffered and streamed off-device upon a detection event, consuming more system resources and power.

One solution to reducing latency is to configure the DNN-based detector to exclude future sensor data from models that are making decisions about the current frame. For example, you can reduce the latency from 3 seconds to 0.25 seconds, and let the sensor rely on longer time spans from the past, rather than future data, to make decisions.

What Can Advanced AI Algorithms Reveal?

Using machine learning and DNNs in PIR sensors not only uncovers much more information about a trigger, it enables the sensors to be paired reliably with other technologies, such as an alerting mechanism that automatically sets off an alarm and contacts the police. DNNs can be used in a wide range of motion sensor applications, such as manufacturing, robot navigation and agriculture, to produce actionable detections and enable higher degrees of automation and on-device intelligence, even in low-power compute environments.