What is the cheapest / easiest way of detecting a person?

Question

I'd like to know if anyone has had success detecting a warm-bodied mammal (ie. Human) using standard off the shelf, inexpensive sensors?

Ideally, I'd like to use an inexpensive sensor or combination of sensors to detect a person within a room and localize that person. I would like the robot to enter a room, detect if a human(s) is/are present and then move to the detected human. The accuracy does not need to be 100%, as cost is more of a factor. I'd like the computational requirements of such a sensor to be such that it can run on an Arduino, although if it's impossible, I'd be willing to utilize something with more horespower, such as a Raspberry Pi or a BeagleBone Black. I have a few thoughts; however, none of them are ideal:

1. PIR Sensor - Can detect movement within a large field of vision (ie. usually 120 degrees or more). Might be the closest thing to a "human" detector that I'm aware of; however, it requires movement and localizing/triangulating where a person is would be very difficult (impossible?) with such a large field of vision.
2. Ultrasound - Can detect objects with good precision. Has a much narrower field of view; however, is unable to differentiate between a static non-living object and a human.
3. IR detectors - (ie. Sharp range sensors) Can again detect objects with great precision, very narrow field of view; however, it is again unable to differentiate objects.
4. Webcam + OpenCV - Possibly use face detection to detect human(s) in a room. This may be the best option; however, OpenCV is computationally expensive and would require much more than an arduino to run. Even on a Raspberry Pi, it can be slow.
5. Kinect - Using the feature detection capabilities of Kinect, it would be relatively easy to identify humans in an area; however, the Kinect is too expensive and I would not consider it a "cheap" solution.

Perhaps someone is aware of a inexpensive "heat-detector" tuned to body heat and/or has had success with some combination of (#1-4) above and would like to share their results?

Answer 1

A combination of a passive infrared detector (PIR) and sonar range finder (SRF) should do the trick.

What has worked well for me previously (not finding humans but very similar) was to have two PIRs on the left and right sides pointed so they have a little bit of overlap in the middle.
You can then figure out if the human is to the left, right or in front (when both are on). You basically then stack this on top of the SRF which will tell you range etc. It is a bit dirty and you have to make some assumptions, but it works well for it's simplicity.

The pseudo code for the 2 PIRs could be something as dead simple as:

amount = 60; //degrees
while (notCloseEnough)
{
  if (bothActive)
    forward;
  else 
  {
    if (leftActive)  
      turnLeftByAmount(amount);
    else
      turnRightByAmount(amount);
    amount = amount - 5;

    //recalibrate
    if (amount <= 0)
      amount = 60;
  }

  checkIfCloseEnough();
}

The idea is that you turn a lot to one side (60 degrees) if you see something in that area. If they are not in front of you after the turn, turn a little bit less to the side which you are seeing them. Keep repeating and narrowing the amount of turn until they are in front of you, then forward. Remember that you do not turn as much (reset the angle) once they are in front because the will not move 'out of scope' as quick.

I was genuinely amazed by how good this algorithm actually works (we used it for automated chase toys and had to slow/dumb it down because it would beat/catch a human controlled robot too easily).

Both sensors are available from Pololu (no affiliation):

Passive Infrared Detector
Sonar Range Finder

Answer 2

A more recent sensor type that can be used are MEMS based temperature arrays by Omron (D6T range) or Excelitas (DigiPile). These, as opposed to PIR elements, measure absolute temperatures and thereby allow to distinguish between background and foreground temperatures and detect movement and static presences of temperature sources.

Answer 3

a capacitive sensor could work, it's really cheap to make, just aluminum foil and a few resistors, it can detect flesh but i'm not sure if iy doesn't detect anything but flesh, you can use 3 to triangulate

Answer 4

I can't say whether this is easiest, but conceivably you could use the Eulerian Video Magnification library to detect the pulse of a person.

In that case, you would be looking for a fluctuation in the video that matched the expected range of human pulses. You would also need a clear image of a body part that exhibited the visible pulse.

There has also been some work (example 1, example 2) exploring hardware-based face detection. Digital cameras from a few years ago had this capability, which was essentially a highly optimized neural network designed to say "does this square contain a face or not"... then you just iterate over a set of predefined squares in the captured image.

Answer 5

I tried using PIR but it has problems with handling and delays. It is not an efficient choice for human detection to be honest. You can use Capacitive Sensing Technique as it is the cheapest and easiest way for human detection(A smart choice) and it is less complex too. You can make a sensor for yourself at very low cost and it is good for small projects. I have used one in my "Human Detection Robot" Project. You can watch my video at: Capacitive-Based Human Detection

Answer 6

A non-bare-metal solution that will probably become increasingly popular over the next few years would be to offload your heavy data processing task (e.g recognizing a human in an image) to a Cloud service. That's assuming your device is connected to the internet. Here is an example with the Raspberry Pi and the Google Cloud Vision API : https://www.dexterindustries.com/howto/use-google-cloud-vision-on-the-raspberry-pi/. Note that it requires a subscription to google cloud past a trial period, but some other cloud vision APIs (Amazon, Microsoft Azure, ... ?) may even offer their services for free if you submit less than N requests per month to their servers.

Yet another solution for heavy data processing on small platforms would be to offload the work to a thumbdrive-like device on your robot with a dedicated processor unit for running already-trained machine learning models (e.g Movidius Neural Compute Stick with the Raspberry PI: https://medium.com/deep-learning-turkey/a-brief-guide-to-intel-movidius-neural-compute-stick-with-raspberry-pi-3-f60bf7683d40 ). This works offline too. They're still a bit expensive for hobby projects but I expect their cost will go down like everything.