With the spread of the COVID-19 pandemic, countries worldwide began to implement non-pharmaceutical and pharmaceutical policy interventions to mitigate the transmission of the virus SARS-CoV-2. Our study focuses on non-pharmaceutical policy interventions. About one and a half years after COVID-19 was declared a global pandemic, several studies on non-pharmaceutical policy interventions and their effectiveness in reducing key pandemic indicators, such as growth rate, incidence, or the R-indicator, have been published. Here, we summarize some of the key findings of several recent studies. Although the studies differed in data and methods used, it was overall shown that lockdown periods, school/university closures, closing workspaces, physical distancing, and closing (non-essential) businesses were highly effective non-pharmaceutical policy interventions reducing the virus transmission. Some inconsistent findings are reported for public transport closure, testing, contact tracing, and (international) travel restrictions. The effectiveness of, e.g. public information campaigns, restrictions on internal movement, or restrictions on public transportation was not conclusively clarified [17–25]. Currently, the most appropriate tool to quantify the severity of non-pharmaceutical policy interventions seems to be the Oxford Covid-19-Government Response Tracker [26, 27], which we will use in our study (for details, see section Policy intervention data). To study how populations have reacted to these interventions during the pandemic, population surveys or data generated through mobile phones have often been used. In the following, the challenges and limitations of these data sources will be outlined and compared to the sensor-based real-time measurements presented in this paper.

Population surveys were used worldwide to collect data on adopting pandemic-preventing policy interventions [28–33]. Because person-to-person transmission causes the spread of SARS-CoV-2, face-to-face surveys were ineligible [34, 35]. Therefore, web surveys were used because no interviewers are required. Furthermore, web surveys represent low costs, short fieldwork periods, and timeliness of the data [36]. However, the use of (web) surveys has several problems in evaluating policy interventions for pandemic-preventing behaviour. First, reported and actual behaviour might differ due to social desirability [37]. Second, memory errors can lead to a bias in the responses [38]. Third, the time gap between interviews and publication hinders a quick assessment of policy interventions. Fourth, the data collection for continuous monitoring is costly [39]. Fifth, web surveys yield biased results for behavioural variables regarding health [40–43]. Finally, most web surveys are based upon non-probability samples and thus have severe restrictions regarding population inference [44, 45]. This aspect of web surveys can be considered as their main issue.

Data generated through mobile phones allows the real-time assessment of actual behaviour and does not rely on people’s memory. Several studies based upon GPS, mobile phone position data, and specific smartphone applications have been conducted during the COVID-19 pandemic. For example, a number of studies used real-time mobile phone position data to study changes in human mobility behaviour in the USA, China, and Europe [28, 46, 47]. A further study based upon volunteered geographic information (app data), assessed worldwide mobility [48]. Data from companies such as Google or Baidu was used to study human mobility and the effects of policy measures [49, 50]. However, the data used in such studies suffers from self-selection procedures. The required device has to be owned, the application needs to be known by the owner, no privacy concerns should exist, the willingness to participate is a requirement, owners have to activate the required service of the device/application, and the device needs to be carried continuously. It is unlikely that these consecutive steps will yield a random selection of the population [51]. Hence, using data generated through mobile phones (either location-based or app-generated) to describe changes in mobility or evaluate the effects of policy interventions, will exclude specific subgroups of the population. Additional problems using GPS-generated data are the accuracy and completeness of the data. GPS devices cause problems due to intended or unintended switch-off, delays due to standby mode, battery issues, or the device not being carried [52–54]. Accordingly, the usefulness of smartphone-based technologies to mitigate and monitor the pandemic is still debated [51, 55, 56].

The spatial mobility data measured by the sensors used in this study include location-based measurements of pedestrian flows in different cities and has received little attention in the scientific literature during the COVID-19 pandemic (e.g. [57]). The recorded data of such systems has several advantages compared to those previously described. First, in contrast to population surveys, these sensors provide objective in situ measurements of the population’s behaviour, overcoming problems due to survey respondents’ social desirability and memory errors. Furthermore, the delay between recording and reporting results and costs for continuous monitoring are negligible when analysis can be done promptly. Second, in contrast to data generated through mobile phones, an important difference is that these sensors measure only whether public places or locations are still visited, i.e. they do not measure who has visited them. As such, data privacy and protection, which were not always adequately addressed during the pandemic [58], are no issue when using such data. Coverage and selectivity issues in smartphone usage among the population or the required use of specific smartphone applications can also be neglected. Finally, data obtained through such objective in situ measurement systems offers the potential for official statistics output. More specifically, given the number of advantages mentioned and the results provided by this paper, this type of technology and the data collected could be more widely used by NSIs and policymakers, for example, future crisis management or policy evaluation. Until now, only a few applications of such data in official statistics can be found, such as the recording of traffic with traffic loop data [8, 59] or the use of real-time indicators based upon publicly available streamed traffic camera images to monitor busyness during a lockdown [60].

The data of sensors measuring pedestrian flows in metropolitan areas is provided by the company hystreet.com GmbH. The data collection started on 01.05.2018 at 27 locations in Germany. The sensor network has been continuously expanded, and currently, sensors are installed in 74 cities at 152 locations in four European countries (Germany, Austria, Switzerland, and the Netherlands). Sensors were installed strategically at these locations from a real estate perspective. A real estate perspective is to determine the possible turnover potential of a location, assess the location quality of the real estate object, and evaluate the attractiveness of cities. As a result, the sensors are located at centres of metropolitan areas on economically relevant and highly frequented streets and places. According to the definition of the German Federal Office for Construction and Regional Planning, 45 out of 49 analysed cities can be considered metropolitan areas (≥100,000 inhabitants). The remaining four cities are medium-sized cities (≥20,000 up to 100,000) [61]. The different sizes of the cities are not taken into account in the analysis, given the majority belong to the same category (metropolitan area), and given that Brocker and Klingwort showed that the developments in pedestrian flows were comparable across cities [62].

The sensors are attached to facades and generate an invisible and eye-safe quadruple light curtain to measure the pedestrian flows constantly. Hence, pedestrians are not aware of being recorded and cannot consciously avoid the recording. When pedestrians cross the light curtain, a record is made. An accuracy of 99% can be achieved up to a flow rate of approximately 500 persons per minute. Persons crossing the light curtain multiple times are recorded repeatedly. Hence, the total count per day might not be a count of unique persons. For example, persons on their route to, e.g. essential shopping and back, might cross the light curtain twice if they take the same route. However, it is reasonable to assume that such scenarios do not systematically upward bias the pedestrian counts. Only if persons cross the light curtain, in the most extreme case, in the form of a curved line, a form of systematic measurement error will occur. This scenario seems unlikely as people are not aware of the light curtain. Measurement errors can occur if the laser is blocked (e.g. by scaffolding, cranes, or treetops), and in case of energy failure, no measurements are made [63]. When these types of measurement errors occurred, data was excluded from the analysis, as we will explain in the following paragraph.

Data from 01.01.2019 to 18.04.2021 of German sensor locations will be used for the analysis. Sensors with either considerable down-times (about one year) or down-times of at least about three weeks in the periods of interest were excluded from the analysis. The resulting dataset consists of 49 cities with 100 sensor locations. The cities have a minimum of about 25,000 inhabitants, on average about 400,000, and a maximum of 3.7 million (Berlin, capital city of Germany). Figure 1 shows the considered cities and indicates the number of sensors per city. On average, there are two sensor locations and a maximum of nine per city. 14 out of 16 federal states have at least one sensor location because one has no sensors installed and the second is omitted due to the selection criteria.

On 11^th March 2020, the World Health Organization (WHO) declared COVID-19 a pandemic [64]. After that, policy interventions were enacted, extended, relaxed, or reintroduced in Germany at 15 points in time (in the period considered in this paper). Since the interventions aimed to reduce the number of social contacts, it is reasonable to assume that closing schools, universities, (non-essential) businesses and restricting public gatherings affect pedestrian flows at public places. To quantify the stringency of the measures, we used the ‘stringency index’ from the Oxford Covid-19-Government Response Tracker (OxCGRT) project [26, 27]. This project systematically records policy measures in over 180 countries that governments have implemented since 1^st January 2020, to address the pandemic. The stringency index has been calculated since 21^st January 2020, and consists of 9 indicators recording the severity of lockdown style policies (school closures, workplace closing, cancel public events, restrictions on gatherings, public transportation, stay at home order, restrictions on internal movement, international travel controls, public information campaigns) that primarily restrict people’s behavior [27]. Most of these policies were considered highly effective in mitigating the transmission of the virus SARS-CoV-2 (see section “Background”). The stringency index is calculated using all ordinal containment and closure policy indicators, plus an indicator recording public information campaigns. The index ranges from 0-100 (100 = strictest). Table 1 shows the dates when policy interventions were enacted, describes the interventions, and shows the stringency index on that specific day.

First, we focus on the period from 01.01.2020 to 18.04.2021, which contains 15 points in time with policy interventions (see Table 1). For this period, policy interventions were linked to the pedestrian flow data using the date as a unique identifier. This allows observing potential changes in pedestrian flows before or after interventions. Within this period, 0.1% of the observed data contained zero counts. The sensors store a zero-count in case of a measurement error and for the actual absence of individuals. However, zero-counts on a given day are implausible, even in the case of a strict lockdown or an evening/night curfew. Therefore, a conditional median imputation was applied to correct the zero counts.

Second, the period from 01.01.2019 to 31.12.2020 is considered to quantify differences in pedestrian counts between the pandemic (01.01.2020 to 31.12.2020) and the pre-pandemic year (01.01.2019 to 31.12.2019). A complete time series of sensor data is required to calculate the relative difference between 2020 and 2019. As the sensors were installed consecutively over time, the complete time series for 2020 and 2019 were only available for 20 cities and 40 locations. Within this period (01.01.2019 to 31.12.2020), 0.6% of the observed data contained zero counts. These were corrected using the method explained above. Due to the weekday shifts of the same date between years, leap year, public holidays on different weekdays, the pedestrian counts of each location were summed per corresponding city and week. No decomposition of the time series into its systematic and unsystematic components was done. Therefore, we cannot tell whether the pedestrian counts might depend on seasonal influences. The relative difference between 2020 and 2019 is calculated using the 2019 data as the benchmark. Accordingly, a positive relative difference indicates the counts in 2020 being larger than in 2019, and vice versa, a negative relative difference indicates a lower count in 2020 than in 2019.

Figure 2 shows the pedestrian counts of all considered cities and locations and the average daily count in the period from 01.01.2020 to 18.04.2021. In general, the counts from Mondays to Thursdays are comparable, rise on Fridays, peak on Saturdays, and reach the lowest level on Sundays. In Germany, there are peculiarities regarding the opening of stores on weekends. On Sundays, the drop in pedestrian counts is due to almost exclusively to all shops, including essential businesses, being closed. Exceptions include train stations, gas stations, bakeries, restaurants, cafes, and pubs. However, the latter three would have been closed in the pandemic year during a lockdown anyway.

Until early March 2020, a similar and repeating pattern in the pedestrian counts is observed. Shortly after the WHO declaration on March 11th, 2020, pedestrian counts started to decrease. Here, at the same time, the stringency index started to increase. On March 13th, 2020, the first national policy interventions (school closings) were enacted.

The pedestrian counts decreased further before the second policy intervention on 22^nd March 2020, when non-essential businesses closed. During the following period, pedestrian counts remained low, although some days show high pedestrian counts at a few locations. Throughout this period, the stringency index increased further. The pedestrian counts started to increase slightly before the interventions were lifted on 4^th and 6^th May 2020. This finding suggests that behaviour already changed before the policy interventions were lifted. Thereafter, pedestrian counts increased and remained stable during the summer months. Here, the stringency index indicates lower severity of policy measures. After the introduced interventions on 2^nd November 2020, almost no decrease in pedestrian counts was observed. A reasonable explanation might be that wholesale/retail businesses remained open, which are likely to be located in the areas around the sensor locations and still attracted customers. With the introduced interventions on 14^th and 16^th December 2020, a substantial decrease in pedestrian counts was observed. This development is accompanied by a steep increase in the stringency index. During these interventions and their threefold extension, the pedestrian counts remained at a low level and the stringency at a high level. An increase after the stepwise re-opening of schools on 4^th January 2021, could not be observed. From 1^st March 2021, with gradual openings and relaxations, slight increases in pedestrian counts were recorded.

Figure 3 shows the weekly relative difference in the pedestrian counts between the pandemic (2020) and pre-pandemic (2019) year. On average, the pedestrian counts were lower in the first week of 2020 compared to 2019 (-1%). Between the second week up to the sixth week (beginning of February), the average pedestrian counts were larger in 2020 than in 2019 (from 2% up to 13%). From the sixth week, the average pedestrian counts fall below the level of 2019. The pedestrian counts are beginning to decrease in 2020 before the WHO declaration and first policy interventions in week 11 compared to the previous year.

The most considerable difference was observed in week 13 (-85%), the second week after the first restricting policy interventions. The difference remained stable between weeks 13 to 15 and steadily decreased from week 16 (second half of April). Hence, pedestrian counts developed towards the previous year’s count about three weeks before the policy interventions were relaxed in week 19. In the following weeks, the average difference converged further to the level of 2019 but remained lower. The difference was positive at some specific locations, i.e. the pedestrian counts were larger at these locations in 2020. In week 30 (end of July), the average difference was about zero. Until week 43, there was some variation in the relative difference (from -14% up to -23%). Starting in week 43 (second half of October), the average difference started to increase again. Hence, before the policy interventions in November and December 2020, the pedestrian counts started to decline again compared to 2019. By the end of 2020 (weeks 51 to 53), the relative difference was comparable to weeks 13 to 16 (-74% to -83%). A reasonable explanation for this observation might be the same type of policy interventions (see Table 1).