New ways to count pedestrians

Experiences from a pedestrian counting experiment with Placemeter

As soon as a pedestrian or cycling planning project appears in the pipeline, you have to think about counting methods. Knowledge of flows and densities is essential for an intelligent and safe infrastructure design and before and after evaluations. The traditional method to perform such evaluation is to conduct manual counts: a labour intensive and therefore costly endeavour. Emerging technologies promise to cut down on counting cost. The folks at Alta Planning did a great job and published a White Paper which covers key emerging technologies in addition to the existing NCHRP 797 guidebook on pedestrian and bicycle volume data collection.

For our ongoing research project Engaging Active Mobility at the Future Cities Laboratory in Singapore, we are interested in testing the viability of some of those new technologies to count pedestrians and cyclists.  Four our first test, we used sensor products of Placemeter, a startup founded in 2012 in New York. Placemeter positions itself  as an ‘urban intelligence platform’, which ‘ingests any kind of video to analyse pedestrian and vehicular movement, revealing hidden patterns and strategic opportunities’, to help ‘build stronger businesses, efficient cities, and innovative neighbourhoods worldwide.’

Placemeter’s counting technology

Placemeter currently supports two types of counting devices: the Placemeter Sensor, which contains a camera and on-board processing unit, and an off-the-shelf IP camera (see Figure 1). In the case of the Placemeter Sensor, the video data is directly processed in the sensor, while the video stream of the IP camera is broadcasted to the Placemeter servers which run the algorithms that extract the count information. In both cases, the algorithms are able to identify counts by the direction of movement across used-defined measurement points. However, currently the Placemeter products do not allow to differentiate whether a pedestrian, cyclist and motorized vehicle crossed a measurement line.

fig_1
Figure 1: Placemeter Sensor (left) and IP camera (right)

Installation Set-up

We ordered both type of Placemeter Sensor products in October 2015 at the price of USD 90$ each. Delivery to Singapore was another 30$ each. While Placemeter promises that measuring with your very first sensor will be free forever, for any additional sensor they charge 100$ per measurement point and month, but allow you free counting during the first month.

The Placemeter Sensor arrived within a few days in beautifully custom-designed cardboard box while the IP camera by DLink got shipped in the standard wrapping.

The technical equipment is principally very easy to install. Nevertheless, some issues appeared when connecting the Placemeter Sensor to our WLAN. This step is necessary to connect the sensor to to Placemeter’s servers and to specify the measurement points through Placemeter’s web-based interface. We then tried to narrow down the problem by setting it up within other WLANs, but without success. Raising the issue to the helpful Placemeter customer care, they offered to replace the sensor at free cost and to cover any shipping expenses.

The setup with the replacement sensor was then straightforward and did not take more than 10 minutes. After connecting, the live stream of the sensor is displayed online and user-defined measurement points can be drawn as a line direct on the picture.

Besides access to WLAN, both the IP camera and Placemeter Sensor require electric power supply. The  enclosed power cable for the Placemeter Sensor is about 5m long, the standard cable delivered with the DLink IP camera is about 2m. While you obviously always can use a cord extension to extend your setup range, it also means that you need to have access to a power outlet near your count location.

First results

IP Camera

While the Placemeter Sensor was being replaced, we started to install our IP camera and decided to place it in front of our main meeting room, the ValueLab. Figure 2 shows the indoor setting and the three measurement points (screenlines) placed in order to count people coming from left and right and entering the meeting room. We can see that our office is brightly lit with artificial light, but also bright daylight as is typical for equatorial regions. However, the combination of the dark floor, white walls and bright windows plus some obstructing furniture create a situation that is characterised by high contrasts.

fig_2
Figure 2: Screen shot of the IP camera live stream and fixation on the ceiling

The ValueLab, also serves as a venue for public lectures. These lectures offer the perfect opportunity to conduct manual counting and compare the results with the counts produced by Placemeter. Figure 3 shows the number of people counted at the door (horizontal screenline) with the IP camera as well as the manual counts. In order to simplify the manual counting, we simply counted the number of people who attended a lecture. By multiplying this number by a factor of two, we obtain a minimum value of the number of people who traversed the measuring point at some point during the measurement period.
In other words, the real value of the manual counts should be higher, taking into consideration that some people would enter and exit the room few times (e. g. to take a phone call or so). Nevertheless, in almost all cases the manually counted ‘minimum value’ was higher than the amount of participants recorded by the IP camera.

Figure 3: Comparison of results from IP camera and manual counts

Another test was to compare the monthly data obtained from Placemeter for the IP camera. Since the ValueLab has a single entry point, we just compared the number of persons entering and leaving the room. During the period of 30 days, the IP camera counts in average 13% more people entering than leaving the room. We think this is because the arrival process is different from the departure process. People trickle in before a lecture, and rush out simultaneously to grab a coffee, challenging the Placemeter counting logarithms.
We also suppose that the IP camera has problems to count pedestrian just before or after a door, since in this case the camera angle is reduced and only one side is visible.

Placemeter Sensor

Not being truly convinced by the counting performance using the IP camera, we also tested the newly arrived replacement Placemeter Sensor in this challenging indoor setting. However, results showed similarly inconsistent patterns. Apparently, the the high contrast setting is beyond the operational limits of Placemeter’s video processing count algorithms.
Disappointed to not being able to automatically count the popularity of our main meeting room, but not disheartened with the potential of the technology altogether, we set out to give the Placemeter Sensor a second chance in a more conducive environment. The setting for the second experiment is the entrance to one of the CREATE buildings. The video frame is characterised by a rather homogenous contrast as shown in Figure 4.

Figure 4: Screen shot of the Placemeter Sensor live stream and list of measurement points

The measurement points (which actually are lines) were drawn to form a measurement square which means that people entering the square should exit it again within a few seconds. This allow us to do a quick accuracy test of the Placemeter Sensor.
Figure 5 shows an overview of the manual count results and the data delivered by the Placemeter Sensor. Different to the outcome of the first experiment, the manual counts are exact (no minimum values). Manual counting was performed for the duration of one hour at three different days. Except for the measurement points ‘Door_Create_In’ and ‘Door_Create_Out’ the results match pretty well with the Placemeter Sensor outputs.
In general, the results from the horizontal measurement points are less accurate then the data from vertical ones. Thus, we suppose that the video processing algorithms can recognise pedestrians better on vertical measurement points, not least because of a favourable angle between pedestrian flow and measurement point.

Figure 5: Results of manual and automated counts on an outdoor setting
Figure 5: Results of manual and automated counts on an outdoor setting

Similar to the IP camera, we made also in this case a second sanity check by comparing the number of people entering and leaving the area between the four measurement lines; over a period of two weeks, the Placemeter Sensor counted in average 10% more people entering this area than leaving this area. Apparently, we have some sort of Bermuda square here 😉 !

Conclusion

In general, with both Placemeter Sensor and IP camera, the detected pedestrian count is lower than the actual count. With our small sample, we made the following observations:

  • Both the IP camera and Placemeter Sensor seem to have problems to deal with groups of persons.
  • Especially in the cases where people are walking through a door, the sensors seem to be less accurate than manual counting.
  • The accuracy of vertical screenlines is better than horizontal measurements screenlines.
  • Likewise, both sensors have problems in indoor settings consisting of furniture, high colour contrasts and changing lighting conditions.

Even though we see a demonstrable utility in the conceptual idea of Placemeter, we learned that the field of application of Placemeter Sensors seems to be limited by the visual setting and the requirement of having access to a power outlet and a WLAN. For peak hour pedestrian flows and public transport measurements, a shorter time interval might be necessary. However, if the situation where you want to count fulfils those criteria, you should definitely consider Placemeter’s products for your counting project as the setup is straightforward and data can be easily collected for long periods of time at marginal additional cost.

Leave a Reply

Your email address will not be published.