Windmill Data Acquisition Software Logo
22 April 2020

Windmill Software
Data Acquisition Intelligence
Call now: +44 (0)161 834 6688

Solving the five challenges of Human Sensing

Researchers have identified five technology challenges in human sensing systems.

Thiago Teixeira, Gershon Dublon and Andreas Savvide of Yale University and Massachusetts Institute of Technology addressed the increasingly common requirement of computer systems to extract information about the people present in an environment. They pin-pointed five requirements that the most accurate systems need to meet.

Human sensing covers not just whether any people are present, but how many and where they are.

There are several methods of human sensing, including those using infra-red sensors, cameras, pressure mats, mobile phone signals, break beams, movement sensors or turnstiles.

1. Environmental Variations

Unexpected or sudden changes in environmental conditions are common sources of errors in some real-world scenarios. Radar signals, for instance, can be dampened by rain or fog. Passive infra-red sensors are often triggered by heat currents flowing through HVAC (heating, ventilation and air conditioning) systems. A large portion of the computer vision literature is aimed at dealing with variations in lighting, shadows and so forth.

Some of the most accurate human sensing systems are those employing video technology with software algorithms to differentiate people from the background. Users can configure the software for each individual sensing station: adjusting for dark shadows and changes in lighting. Tests show that video technology achieves over 98% accuracy in human sensing.

2. Similarity to Background Signal

Clearly, separating a person from the background signal is a core requirement for human-sensing the scientists write. The latest video people sensing systems freeze the background at the moment a person enters and easily distinguishes the moving person from her surroundings. Users can set in software how long a person can stand perfectly still before the system sees them as part of the background.

In other domains, such as with ranging sensors (radars, sonars), the presence of unwanted signals with the correct frequency spectrum or timing characteristics can often fool the system into producing phantom detections.

Passive thermal sensors have problems differentiating people in warm environments. The sensor works by monitoring the temperature difference between a person and the background. If the temperature of the background and the person are similar then the sensor doesn't "see" the person. Also, if a person stands still she becomes part of the background.

3. Appearance Variability and Unpredictability

People look different and wear a vast assortment of different types of clothes and hats, they push trollies and pushchairs, and carry back-packs and hand-bags.

This at first looks like a problem for video systems, but is solved by the software converting an individual to a "blob".

In a tricky situation - outdoors with people pushing prams - the computer vision system simplifies the picture and accurately distinguishes people, converting them to "blobs". Video courtesy Retail Sensing

Researchers have identified five technology challenges in human sensing systems.

Thiago Teixeira, Gershon Dublon and Andreas Savvide of Yale University and Massachusetts Institute of Technology addressed the increasingly common requirement of computer systems to extract information about the people present in an environment. They pin-pointed five requirements that the most accurate systems need to meet.

Human sensing covers not just whether any people are present, but how many and where they are.

There are several methods of human sensing, including those using infra-red sensors, cameras, pressure mats, break beams, movement sensors or turnstiles.

4. Similarity to Other People

Some tracking systems use identifying features of a person to track them - presenting a challenge if people are all wearing similar clothes or uniforms. Video sensing technology follows the "blob" that has been identified as a person so it doesn't matter if everyone looks alike. Each blob is seen as unique and distinguished from the others, even in crowded situations.

5. Active Deception

The researchers' final point is when a human sensing system may be deliberately debilitated, perhaps by people walking slowly to fool motion sensors, covering a break beam with their hands or turning the lights off to fool the cameras. A cctv system lets users remotely play back the video over the internet to see why counting has suddenly stopped and rectify the problem.

And the best human sensing systems?

The researchers conclude that the best system is computer vision, saying "Computer vision is far ahead from other instrumented modalities not only with respect to spatial-resolution and precision metrics, but also in terms of having the most field-tested solutions".

Further Reading

You can read the research at A Survey of Human-Sensing: Methods for Detecting Presence, Count, Location, Track, and Identity T TEIXEIRA, G DUBLON, A SAVVIDES ENALAB technical report

How a CCTV People Counting System Works, Retail Sensing