Due to the biological unpreparedness in relation to the toxicity of the urban air, humans can not easily recognize the threat that surrounds them in their everyday life.
A simple worn device with a power source, camera and minimal actuators could complement the human senses by sending haptically recognisable warning impulses to the user, guiding them through the city on safer and more enjoyable pathways. Pictures taken of the surrounding air, analysed through artificial intelligence (AI), could provide an approximative air quality factor, and furthermore determine possible routes through the surroundings. This technology would act as a “new sense”, supplementing the human ones. It would be a demonstration of possible human-computer-interaction and extension of human capacities through technology enlarging the ability to make conscious decisions in everyday life.
We imagine the wearable device as a headband that expresses the technology it incorporates. Designed in a conspicuous way, its wearing alone would become a factor for raising the awareness about damaging factors in the air, attracting attention about today's air quality issue to the vast majority of citizens. A second layer of information is the existence of a publicly available map providing all collected sensor data, constantly updated with real time measurements. Through training a machine-learning-model to recognize air quality from pictures, anyone with a camera and an internet connection could participate in sourcing such data, creating social involvement to provide both constant data, more recognition and understanding concerning air quality.
Furthermore, the permanent collection and evaluation of air data, driven by artificial intelligence, enables an extremely detailed analysis of the impact the built environment has regarding the air quality. Thus, these new insights can make an important difference in unpreceded healthier concepts of urban planning.
Human understanding of the universe goes as far as recognizing the improbability of what life on Earth turned out to be today, even the understanding of our own planet has always been limited by the technological limitations of each era. As an invisible layer enveloping the entire planet, the earth’s atmosphere is a collection of gasses held together by gravity that is essential to the formation and flourishing of life, of intelligence and ultimately of humanity. And yet, its specificities, its subtle changes, its alterations produced by both time and the unfolding of life, its destruction through unbalanced and exploitative activities, are all impossible to accurately ascertain through human senses alone.
In recent years, not only the technology has reached a point where it can clearly establish, through analysis of localized atmospheric properties, its effects and repercussions over society. An individual’s level of interconnection, of participation into a network of sharing data and information, also establishes potential for real time and extremely specific evaluations of the surrounding atmosphere with a meaningful impact over individual living spaces, no matter how temporary that is defined.
The introduction of artificial intelligence (AI) into analytical processes is almost metaphoric, as the processing and analysis of data is done digitally because of human limitations. Yet it is the overall goal of AI to develop a more human-like approach to learning, to simulate a brain, it’s neurons. Hereby it is important to distinguish a clear additive approach towards the complete information humans could have access to in a process of enhancing their senses with conclusions drawn from factors they cannot perceive through physical senses alone. The effort of understanding one’s environment and analyzing it has the potential of being reencoded and actuated into supplying additional information through lesser used senses, offering an implicit understanding of qualities of the surrounding atmospheres that could otherwise not even be perceived.
An online database could be created from a large and widely distributed amount of data, with corresponding measurements of air particles, different gas proportions and corresponding imaging and geographic coordinates. This would provide the base of training a neural network to determine air quality from a narrower recording. A combination of simple imaging and information already provided through a smartphone device, in addition to an internet connection, could suffice to establish air quality around the user, and provide haptic feedback, engaging the touch as one of the lesser used senses especially while navigating through urban environments. This process addresses a very important concern of pedestrians in urban spaces: the air they breathe, the toxicity that affects them, that they are not able to perceive without external means. A continuous supplementing of the database through this constant user-based analysis ensuring lasting and constant improving diagnoses of atmospheric qualities, creates a digital mapping of the human environment with the potential of recreating local networks of urban pedestrian movement through healthier, more breathable spaces, providing therefore a positive impact on society as a whole.
In today's context of constant population growth and densification of urban areas, the ambient climate is suffering from radical changes. Compared to the 20th century, with its industrialized cities covered in visible smog, today's industry is still emitting airborne toxins in the air we breathe and their effects on human health are detrimental and long lasting (Smedley, 2019).
Ever since the first publication of the World Health Organization (WHO) Air quality guidelines in 1987, the interest and preoccupation around health impacts of air pollutants has extended from Europe to the entire world, with more and more concerns being raised, especially in dense urban areas in lesser developed countries. The latest update entails a comprehensive explanation of air pollutants, their sources and the health implications they produce.
Primary air pollutants are particles emitted directly into the atmosphere by polluting elements. This category would contain exhaust pipe gas or chimney smoke, but also wind-blown dust. Secondary ones are produced by chemical reactions in the atmosphere, such as ozone forming based on the altitude and the atmospheric composition. This second category is therefore far harder to quantify, or the sources and effects to be tracked down, showcasing the already complex nature of the problem.
Pollutants are also differentiated by type, being either gaseous or of particulate matter, but also by their atmospheric lifetime, which denotes their behavior in the air over time. Some pollutants with short lifetimes can be encountered extremely localized, such as gases emitted by a factory or nitrogen oxides and carbon monoxides produced by traffic, staying highly concentrated around urban areas with crowded roads, their concentration rapidly dropping towards the rural areas. On the other hand, fine particles, such as aerosols of black carbon stemming from burning of fossil fuels or biomass can have lifespans of days, and travel over thousands of kilometre, crossing national boundaries. Carbon dioxide or methane have even longer lifespans . They are associated with greenhouse warming effects, and can persist in the atmosphere for years, distributing all around the globe, prevailing in similar concentrations both around their sources and away from them.
A combination of some or even all these factors can be found anywhere in the world, with sources varying from local transportation to industrial sites to simple households. The density of these sources though has also a major impact on the consistency of the pollution itself. Especially particulate matter and noxious gases with shorter lifespans are concentrated in highly urbanized and densely populated areas. A rapidly growing urban area, particularly in lower income regions, will usually produce a rapid growth of transportation vehicles, and without a well-defined public transportation network, the immediate increase of air pollution is easily anticipated. Megacities around the world, some with over 10 million inhabitants, also display increasing levels of particulate matter, mostly in tandem with population growth, even in some of the most developed countries in the world. (WHO, 2006)
Kolkata, as one of the most populated and polluted cities in the world, makes an excellent case study. With particulate matter, stemming in equal quantities from both transportation and industry, health effects such as the appearance of pulmonary or hematologic afflictions, as well as genotoxic changes that can lead to respiratory or cardiovascular illness or even cancer are tracked. The average Kolkata citizen is shown to be 7 times more affected by air pollution than their rural counterpart, with almost half of the population of the city suffering from respiratory tract infections at the turn of the century. Still, with extremely slow implementation of infrastructure, the education and raising of awareness throughout the general population will still be a major essential step in the remedy of this situation (Spiroska 2011).
While the seriousness of this problem persists, it is not necessarily a direct concern of the general population. While studies show that some degree of particulate matter can be recognized by human vision in certain lighting conditions, it is not necessarily a main factor of discomfort, whereas temperature or even acoustic pollution play a more dominant role (Nikolopoulou, 2009). Some studies even show a mismatch between actual levels of pollution and perceived pollution, with the deciding factor being named as sources of misinformation within the media (Peng, 2019).
Pollution on imperceptible levels can still have lasting short and long-term adverse health effects, and yet, human beings are not equipped to properly perceive it or even react to it. Our senses have been developed throughout millennia to navigate physical environments, to use our senses of sight and hearing to find our way, and our current environments have been designed accordingly. We hear incoming traffic, we have light coordinating our movement, we design specific walking, riding and driving areas, but do not take in consideration the effects these layering’s have. How hazardous the immediate exhaust on the road is affecting us, as the effects are not immediately perceivable, and can not be deduced by the senses we possess.
Pedestrians taking a stroll through the city and enjoying the view are guided safely on their way, led by their own senses, perceiving possible dangers with their sight, hearing, or even smell. Still, these senses can only guide them as far as the threats are recognisable, leaving them vulnerable to the invisible, odorless, and tasteless particles in the air. Based on this missing biological preparedness, a simple worn device could complement the human senses by sending haptically recognisable warning impulses to the user, guiding them through the city on safer and more enjoyable pathways. We imagine the wearable device as a headband that expresses the technology it incorporates. Thus, designed in a conspicuous way, its wearing alone would become a factor for raising the awareness about damaging factors in the air, attracting attention about today's air quality issue to the vast majority of citizens.
Whereas the individual awareness concerning air pollution is still very low, resulting from its invisibility and its very much delayed effects on nature, the general societal and political attention to pollution is already attached to a great importance (Liao, 2015). Different, more holistic approaches are investigated. As one reaction to the different aspects and impacts on the air quality, a combined unit, the “Air Quality Index” (AQI) was introduced, simplifying the readability of different hazardous gases and particles in the air.
Although the bad impact of air pollution is already in the common mindset, a missing feedback on a personal level disables most people to change their circumstances and therefore inhibits individual actions for a better, pollution free atmosphere. But even without direct perception of toxicity, it is up to people themselves to find and create a suitable environment for themselves and others.
For more technical information please refer to the related GitHub-Documentation: https://github.com/ToxSense/AQI
When working with the amounts of data produced by air pollution studies, an implementation solely focused on the conventional approach of mathematical analysis has multiple disadvantages. The accuracy is limited, the methods are inefficient, and very complex mathematical calculations arise, making it hard to readapt them to new incoming data. Therefore, machine learning is already applied in the context of air pollution today, with various approaches to both data collection and actions taken towards remedying inadequate situations. Examples of such models can be used for short term local predictions of air pollution based on a meteorological forecast, correlating it with local air quality factors within a neural network. Another approach is the use of a localized camera to monitor amounts of particulate matter in the air, estimating it from pictures according to the amount of radiation from the sun. This would set an alarm to warn the people in the vicinity of the dangers of going outside during moments of high pollution. Many different types of machine learning approaches have been implemented, but mostly on specific cities, in cases with a particular set of data over a set period of time. Further steps need to take into consideration more dynamic changes such as wind and atmospheric changes over time, to set up reliable prediction models for future use. (Kang, 2018)
Another interesting implementation of machine learning in a similar context is the evaluation of measures to remedy air quality in Beijing, determining whether the different policies were able to achieve their initial set targets. A random forest model would describe the relationship between an air pollutant and its predictors, such as the date and time, and various meteorological data such as wind speed, wind direction, temperature, pressure, and relative humidity. While some correlation between clean air measures and the reduction of some pollutants can be observed, the implementation of multiple measures over overlapping time periods makes their exact evaluation impossible. Still, such approaches, adapted to more exact initial policy parameters show the potential of use for future policy making for Beijing or other cities. (Vu, 2019)
Not only predictive models can be based on machine learning, but also analytical ones. Yi-Chen Wu et al (2017) propose in their paper the use of mobile microscopy to analyze air quality. They show the potential of implementing machine learning into a cost-effective platform controlled by a phone app. This model could be trained to recognize not only the existing particulate matter amount, but even differentiate between different types, such as dust, pollen or mold, a simple device being able to replace a complex and costly sensor setup. (Wu, 2017)
In this project, pictures taken of the surrounding landscape and analyzed through artificial intelligence (AI), provide an approximate AQI. Within the context of already existing air data, possible routes through the surroundings can be suggested. This technology would act as a “new sense”, supplementing the human ones. It would be a demonstration of possible human-computer-interaction and extension of human capacities through technology. What machine learning can achieve through processing the enormous quantity of data available today is to detect numerous correlations in a second that one person on its own could not observe in a lifetime. These results can receive constant measurement-updates, and refined through emerging technologies, such as machine learning, other data processing tools, microcontrollers and haptic feedback devices, the corollary has to always be communicated by simple haptic impulses. The sense of touch is less used for navigating through the urban environment, so the skin would get activated via vibrations factoring right at the edge of conscious perception, conveying significant information in a concise manner.
Considering the large amount of research invested in haptic feedback in the past decades, it clearly has incredible future potential in a variety of fields. In the medical industry, surgical interventions with minimal invasiveness still require human dexterity. In such cases, haptic feedback might allow a surgeon to not have to operate blindly, providing guiding signals based on otherwise imperceptible factors (Sokhanvar, 2013). Robotic teleoperation, enhancement of prosthetic limbs, alternative communication, development of virtual human interaction in artificial reality platforms, these are just some of the discussed topics in this field (Nisky, 2020).
Such technologies help keep humans in a more secure environment by working remotely, or allow them to feel and operate at capacities beyond their biological potential. Or even enable a new level of long-distance interaction. And this is all done not by imposing a foreign technology into human life, but by taking something inherently natural, the sense of touch, and designing an entire technology around it.
This enhancement of perception depth would exemplify how AI would not necessarily replace human actions, but rather complement and augmentate them. A second layer of information is the existence of a publicly available map providing all collected sensor data, constantly updated with real time measurements. Through training a machine-learning-model to recognize air quality from pictures, anyone with a camera and an internet connection could participate in sourcing such data, creating social involvement to provide both constant data and more recognition and understanding concerning air quality.
The above mentioned problematic environmental factors are still appearing in cities all around the world, thus, the investigation continues in parallel, at many locations on the globe (WHO, 2006). Issuing enormous amounts of available data as well as providing possible comparisons and long term predictions, enables global initiatives to implement measures of reducing air pollution and improving the quality of urban life (BreathLife, 2016). On the one hand, a bottom up approach is to provide information to decide and act on one's own benefit at a personal level. This could enable sufficient active movement to recreate some of the social networks within the city, discovering new, safer spaces for human activity. A top down approach would, on the other hand, entail the efforts of designers and planners, as responsible organs for the evolution of urban landscapes, to act upon this analyzed data and the social movements and predictions it provides. They would effectively change the air quality through a more holistic approach to urban planning, that makes the quality of the local atmosphere accessible in decision making processes.
For the user interaction, this project is built in three parts. The first as a haptic feedback device, for returning the measurement and calculations in a natural way. The second user interface consists in an application to control and interact with the device. Lastly an open map provides all information on a global scale.
The hardware setup of the prototype consists of a symbiosis between a microcontroller (ESP32-CAM) together with a smartphone which is driven by an application, that was specially designed and programmed by the projects team, as well as its internet connection to a server. For an interaction between the smartphone and the microcontroller a bluetooth connection is established.
The headband was not designed to be yet another portable device, just like our smartphones or computers, since these are intentionally used devices you have to actively perform with, relying on an active user input. Instead the headband should act as a naturally implemented human sense that one can understand offhandedly and without questioning the way it renders the feedback information. To achieve this, it is important to rely on an already known sense, which is merely used to its full extent: touch.
Despite being the largest organ of the human body, the skin is barely used in everyday life, since it is mostly covered with clothes , leaving the touchsense mainly to the hands.
With this amount of free space to transmit information to the human body, it is still important to define where exactly on the skin the haptic feedback will occur. The decision to create an AI enabled artifact in form of a headband makes it possible to dispatch a 360° degree feedback to the user that is always relative to their own orientation and direction of sight. Positioned at the head, which already provides a multitude of our sensual perception (smell, taste, sight, hearing), it seemed predestined and natural to add a further “sense” to this agglomeration.
The need for a comfortable, sensible, yet detailed feedback drove us to the usage of body bass exciters. They present a real advantage to other possible actuators, as they can act on a wider pressure range with settings for their tone (frequency) as well as their volume (duty cycle) at a completely personal scale to fit the users preferences. A body bass exciter works similar to an audio speaker, but does not create airborne soundwaves, due to their construction without a resonating membrane. Instead, it uses a mass connected to itself as substitution to the membrane and therefore has a very unique way of creating sound (Visaton, 2010). In its usage inside the headband the body bass exciter is directly connected to the user's head, inducing a light vibration , only sensible for one's own. The frequency used for this haptic perception can be set to a personally pleasant number - eg. custom 200Hz - then the duty cycle is tuned to the final volume of the feedback in relation to the calculated AQI , again with a customizable peak volume by a chosen factor.
Together, frequency and duty cycle, build a “pulse width modular signal” (PWM) sent to the body bass exciter.
For more technical information and the source-code please refer to the related GitHub-Documentation: https://github.com/ToxSense/Prototype
Achieving a better mobility, the processing part for the user device is outsourced to an application on the smartphone. Today nearly everybody has a good performing portable computer in their pocket, thus it is not necessary to overcomplicate the functioning of the headband with additional sensors and computing-power. Neither is it required to build yet another device, which has to be carried by the user besides the headband and that uses many unnecessary resources in their production, further inferring climate damaging actions. The developed application combines several functions, for example the querying of the users GPS location, which is needed to provide the user with direction based feedback. That feedback is built by the applications AI. It takes in a picture and outputs an AQI. Since a single index can not give a directional feedback, it is sent to a server together with the GPS coordinates. The server, also powered by a later described AI, returns multiple AQI-values, set in different directions. From there the air-quality in every direction can be calculated by the headband. To establish a link between the server and the headband a bluetooth connection between the smartphone app and the headband device is initiated. Also the picture from the headband, shot in predefined intervals, is immediately transferred to the application over mentioned bluetooth connection.
For standalone use, without the headband, a direct picture-to-AQI function is implemented within the application.
Therefore the smartphone camera provides the picture for the AI model in a continuous stream and displays the calculated AQI in an instant.
To ensure the users consent, the application has to request permissions for the bluetooth connection, the access to the location, the internet connection and the camera permission.
For more technical information and the source-code please refer to the related GitHub-Documentation: https://github.com/ToxSense/Application
To access the collected data, an open map with the current AQI states is provided at the ToxSense homepage: https://toxsense.de/
This access enables the user to run an inference on the map, making it possible to precisely evaluate AQI values on selected positions without a physical presence and therefore making it available not only as an additional human sense. The inference is run by an AI that takes existing measurements, weather-data and a generated map-section for its prediction. Amongst others, urban planners and architects could correlate the impact of different building shapes on the surrounding air quality. The map always shows updated data from the sensor.community15 and the shared, uploaded data from each headband. The values made by the server AI or provided through the headband device are active up to one hour ensuring the accuracy of the displayed values.
Both the online map and the smartphone act as visualization of the collected data, requesting their information from a back-end database over an API connection. The implemented AI receives values from the smartphone or the open map, calculating its inference and adds the result to the database, making it available for everyone on the open map.
For more technical information and the source-code please refer to the related GitHub-Documentation: https://github.com/ToxSense/Server
To understand how the chronological changes of the air quality values occur, a global view on the topic is inevitable. Only world wide collected data gives the AI the potential to calculate and predict values. Thus, the collection of large quantities of measurements and datasets for the training of the AI is decisive and allows it to evaluate meaningful data for the ToxSense-project. There are numerous websites and possibilities to examine current data, in this case AQIs, but unfortunately not many sites offer easy and open access to their databases.
In this project, the datasets are provided from the sensor.community15, an open and free organization of “data-gatherers”, consisting of mostly private people who contribute to the dataset by setting up their own sensors and sharing their measurements to an open API (database access). Due to this organization-concept the ubiquitous sensors generate vast databases, but since it is done by non-professionals, some of the data may be flawed. Therefore it can not be fully relied on the raw measurements and the datasets have to be filtered before usage.
Besides these provided datasets, selfmade local collection boxes, with sensors and a new concept for measuring the air quality, were conceived. Digital sensors provide measurable data to calculate the AQI of the specified area.
Additionally to the purely mathematical approach, the overall liveliness of the area is to be evaluated. The projected test consists of a luminescent bacterial suspension (Aliivibrio fischeri) that is aerated for one hour with surrounding air. If it proves to contain toxic pollutants, the number of bacteria will diminish, leading to a decreased lighting intensity (Institute for hygiene and environment, Hamburg). This decrease would be captured by a camera and calculated in a static value, which again could be taken for AI training and interpolation. All the sensor data is collected in real-time by a purposely programmed Raspberry Pi 3B+, which connects to an online database to save the data. The management of these boxes was realized through the implementation of the open Telegram API. Via the Telegram-Stream different commands can be transferred to the boxes which instruct the execution of various actions (e.g. start/stop recording, shutdown, …).
Even if a massive amount of data is preferable for the training of an AI, it implies numerous obstacles in handling and preparing the data for said training, because of the limited computing-power on our hands. So a large amount of work had to be invested in improving the efficiency of the programs that handle and process the data for the training of the Artificial Intelligence. The AI working on the smartphone is trained on a database structured with photography and AQI-Label pairs. As an addition to the captured data by the Capture-Box, the open Visionair-dataset18 was implemented providing a vast number of data.
For the AQI to be easily predicted, one AI, as part of the smartphone application, is trained to recognize the toxicity of the surroundings through pictures that are captured by the user's smartphone or headband camera. Additionally a second AI is trained on locational and meteorological data to calculate missing values and precise the inference created by the first AI.
Both AI are trained on a combined model using a Convolutional-Neural-Network for image interpretation and a Multilayer Perceptron for the numerical data and its final regression into one single output value6. The network employs a mathematical operation called convolution. Convolutional networks are a specialized type of neural networks that use convolution in place of general matrix multiplication. (Goodfellow, 2016) Convolutional neural networks are a promising tool for solving the problem of pattern recognition mainly used on images (Valueva, 2020). On the other hand, a Multilayer Perceptron is composed of multiple Layers of simulated neurons, that are being activated (or deactivated) depending on different mathematical functions processing the input data (Rosenblatt, 1958). The data is run through this model multiple times, increasing its accuracy in each step, by rearranging the neurons activation and the connections between the layers, comparing the estimated value with the actual data.
The result is astonishing. On average the maximum deviation of the prediction from the actual value does not exceed 10 AQI-Points. These results and their accuracy inevitably raise the question: “How can the AI recognize the AQI out of a single photograph, where otherwise a multitude of sensors are needed?” In our position, a finite answer to this question is impossible, leaving us to guess the abilities of this AI. The trained model is a black-box, where we are unable to query the decisions of the AI. Only deepend research with many tests could provide evidence of which factors influence the decision of the trained model. For the moment, we can only believe in its functioning. But this way of trusting an AI raises very important concerns. An AI only learns from its predefined database, so the curation of the learning-sets obtains a position of great importance in the working process. Is the database flawed, the results will be as well. In many iterations the AI reproduces and even amplifies systemic problems in the collected data. As one of many examples, the criticized risk assessment algorithms for recidivists “COMPAS” largely employed in the United States of America can be cited. According to a report of ProPublica, the AI labeled “white defendants who re-offended within the next two years mistakenly as low risk almost twice as often as black re-offenders (48 percent vs. 28 percent)” (Larson, 2016). This shows that data sourced from a society where social injustice, misogyny and racial discrimination are a matter of course, is not reliable and can even consolidate these problems. Thus the output of AI should never be taken as absolute truth and its base should always be questioned.
For more technical information and the source-code please refer to the related GitHub-Documentation: https://github.com/ToxSense/AI
Creating a “new human sense” is connected to many different aspects of data gathering, data interpretation and on how to return a suitable and simple haptic feedback for the user. The project succeeds in creating a first working prototype with the chosen resources, but many parts have to be enhanced for a realistic everyday usage. The body bass exciters are directly fastened to the headband amplifying the vibration all over its surface, rendering the direction of the changing feedback indistinguishable. Due to this, a workaround was conceived in updating the software to emit only a single point vibration into the direction with the highest AQI, reducing the consistency of the concept that is to convey a sensible feedback perceivable only in the background.
Further, the magnetic sensor should be replaced, since it interferes with the magnetic field of the body bass exciters and can not return accurate values. As a current workaround, the user heading is provided by the smartphone, putting the bluetooth serial connection on a high workload and requiring it to rotate the smartphone simultaneously to the direction of view. As a last enhancement, the microcontroller, currently the ESP32-CAM, may be replaced with another board, as the concurrent usage of its GPIO ports for PWM channels and the camera, as well as an bluetooth connection, lead to synchronicity problems, like the double usage of channels and GPIO pins which ultimately results in crashing different program parts. Since there is no advanced debug terminal, small issues end in long research, taking time better used to update functions or improve the personalisation, fitting each user's preferences.
Another point is the above mentioned availability of data. The data referring to the AQI can often only be reached as visual output on a multitude of websites, but the raw data not being at open disposition a further step with a view to the project would be an API implementation to provide the new collected data or the AI inferred values in a simple way for other developers.
Connected to this point of data-sharing, it is really important to have a look on the data security concerning each user, as well as the transparency on how the data is handled and shared.
For the transmission of data to the server, which contains a location and, for AI-retraining purposes, an encoded image, it is mostly important to encrypt the data and handle it on the server side, respecting every user's privacy. No personal data has to be exposed to others. Only the AI relies on retaining the data, which is deleted after the training process. Currently the location itself is directly displayed on the open map, permitting the tracking of a specific user. This is not intended and therefore a way of circumventing this data breach has to be found and the functioning has to be changed in an appropriate way.
It is desired that the whole project concludes in an open access to the data and an open source licensed program code. This enables others to deepen or change the project for their own purposes, using the data for specific analysis or simply getting themself in touch with their surrounding air quality and its impact. Only in this way can the project make a common change for cities and production. To achieve this, the project already only relies on data and libraries with open source aspects and for further development it is important to check the combination of licences applied at work produced at university and the further used ressources to guarantee an open access and the possibility to share collected results.
Also, the capture box, whose development has started but currently is not implemented, should be finished to provide own and deepened measurements. Those further values would enhance the whole AI with training on a local level and complementing it with a further factor.
The conspicuous design of the headband is meant as a statement of each user. With wearing it, one shows a commitment for the needed change into a better air quality. For non users this is a signal which forces them to think about their own attitude towards air pollution and their own actions regarding it.
With this common awareness, for users and non users, a completely new behaviour and rethinking can be accomplished.
Our somewhat utopic goals can be resumed to a new awareness about toxicity. The inherent and trivial feeling of air-toxicity in everyday life will include the consciousness about its hazardous effects in our every decision. This freedom gathered, it is in our hands to use the acquired knowledge for a better life and to change existing problems. A general awareness will also change the way decisions are made by the people that shape our environment and our society
Through the ToxSense headband, the users can decide freely how they want to act regarding the possible threat of air-pollution. They could choose whether taking a shorter, more harmful path, or opting for a more enjoyable, less polluted, but possibly longer way. Furthermore, the headband implies active strolling, a movement that is not exclusively directed towards reaching a destination, which could inhibit today's accelerated life.
The direction based recommended routes can also pass through commonly unknown parts of the city.
This discovery of new and healthier places and a flexible adaptation through
constantly updated data could lead to high-quality urban experiences and an improvement in the overall urban liveliness.
Within the assumption of a growing use of ToxSense in the future, many scenarios can be imagined where large-scale change could have its roots in the active choices made at an individual level, aided by the proposed system.
The personal decision to avoid higher risk areas, such as major roadways, could entail a new overall stance towards urbanization, which would coincide with current attitudes and the more human-forward approach to urban design. A complete opposite of some of the car-cities envisioned a century ago by the modernist movement, future cities should consider pedestrian accessibility to different city parts as a primary necessity, based on the new social mindset and therefore scaling infrastructure accordingly.
Within developing urban areas, programmes could be overlapped within smaller urban units, to minimize the need of transportation. Trends such as remote and flexible working could provide the basis in new urban planning initiatives for a decentralization even within extremely dense settlements.
Many already existing initiatives promoting air quality awareness already show promising results. The WHO BreatheLife2030 initiative is sponsoring action in cities all around the world, with diverse approaches developed towards the specific local factors. Yet, they all work towards similar holistic approaches, understanding the different layering of solutions and the positive impact that can be ultimately achieved. (Breathlife, 2016)
A BreatheLife article cites Nathalie Roebbel, Head of Air Quality and Health at the WHO: “If we change the structure and planning of our cities to make it easier for people to use bicycles or walk, this will have an impact not only on the air quality by reducing car use but also encourage people to do more physical activity and thereby have a reduction on obesity. Road accidents could also be reduced.” (Climate & Clean Air Coalition, 2020)
The rapid urban densification in Bengaluru, India brought a decline in air quality with large health risks and a lowering of the quality of life. By first deploying monitoring stations to assess the situation, specifically in regions with most vulnerable populations, such as zones with schools and hospitals, the local community is being informed and made more aware about their specific local pollution levels, enabling them to become more implicated in the overall government policies. (Das, 2020)
In Beijing, a platform was programmed to process big data from distributed monitoring stations as well as flexible sensors deployed on vehicles such as taxis, that travel all around the city. Such monitoring allows for a more detailed model that enables the rapid identification of pollution hotspots. This level of information aids the inspection and reglementation of activities such as construction sites, but also industrial and commercial spots, prone to severe air pollution. In comparison to a randomized approach such checkups improve their probability of being made in relevant spots by a factor of ten. (Whitney, 2021)
In Warsaw, the government policies tackle household heating with solid fuels, but also the implementation of electric public transport networks, and a development of alternative travelling solutions through infrastructure such as bike lanes. To enable movement within the city, policies promote urban design in pedestrian scale, accenting the importance of accessibility, while also having informative campaigns to encourage these more environmentally friendly alternatives. (Warsaw Climate Policy Office, 2020)
Even a concept such as correlating air quality data with specific urban morphologies has been investigated. In the case studies of Antwerp and Gdańsk, Geographical Information Systems (GIS) have been applied to correlate urban spatial units with their impact on air pollution, factoring urban ventilation and human exposure to pollutants. Understanding the values of these factors within different configurations of spaces, and their resilience on different geographical terrain and in different climate conditions could become an essential tool for further urban planning that could consider air quality as an initial factor. (Badach, 2020)
Based on these preexisting examples of urban scale interventions, ToxSense could become the means through which individual concerns and preoccupations around air quality could be brought to the level of local planning and urban design. A planner, designer or architect would not only have access to overall air quality data or even predictions, but could see the very result of individuals acting according to that data, the results of new networks and urban movement, and plan accordingly.
When urban growth is noticed, proper infrastructure has to be planned in advance, limiting the time necessary for a transitioning period, especially in developing countries. In already established settlements, these networks can be updated and optimized by identifying problematic points or gaps, and implementing solutions that might have already had positive outcomes in similar urban areas somewhere else in the world.
Architecture will change too, as the spatial conformation of buildings and building clusters is always evolving, repositioning them towards pedestrian friendly zones, more shielded from short term pollution sources. Urban ventilation and a functioning of the design within the geographical configuration would also become an essential considerent, ameliorating the overall air quality at the scale of the city itself and not just the immediate vicinity.
ToxSense does not only present an opportunity to raise awareness on a very important factor in our everyday lives, namely air quality, but also offers an opportunity to exemplify a possible implementation concept for future technologies. While discussion on the opportunities of developments such as AI vary from promising futures to fear and “end-of-the-world visions”, we propose an approach of technology as a human aid, an extension of human abilities, and in this specific case, of the human senses themselves.
Through the use of a wearable, eye-catching object that should bring interest to its very own role and functioning, a statement towards the exterior should be made. The wearer would become an active factor in the movement towards a better, healthier and less risky life within the urban areas, enabling some degree of public recognition of the problem, as well as a simple implementation of problem-solving at the level of the individual.
At the same time, the functioning of the object shows restraint and understanding of human perceptions, and how the feedback of a man-made system can subtly inform conscious human behavior. The simplification, through machine learning, of a vast amount of data into easy to understand signals enables this subtlety. The haptic signals are meant to become a background element, something that one could both concentrate on, or disregard, the device not having any direct impact on human autonomy. This intentional passiveness of the feedback system should help it become more a channel of perception than any set decision making process.
Still, the decision-making process should adapt to a new layer of received information, without it superseding the existing processes. As humans have already adapted to their environments, keeping themselves safe from all kinds of possible dangers, they have the ability to learn to adapt to this perception of air quality as well.
And just as humans have themselves adapted, they have also adapted their environments to their own needs, and should be able to factor air quality into any further developments, as long as they are aware of it. The vast databases that would result not only from this project, but from studies worldwide and overall concern on quality of life within cities, can become part of the basis of any future urban planning initiatives, on both local and even global scale. Learning about our own atmosphere can change the way we live in the foreseeable future.
Air Protection and Climate Policy Office, Department of Air Quality and Education Monitoring in Warsaw (2020), Warsaw’s holistic approach to reduce air pollution, https://breathelife2030.org/news/warsaws-holistic-approach-reduce-air-pollution/ Accessed 2021/03/16
Badach, Joanna; Voordeckers, Dimitri; Nyka, Lucyna; van Acker, Maarten (2020): A framework for Air Quality Management Zones - Useful GIS-based tool for urban planning: Case studies in Antwerp and Gdańsk. In Building and Environment 174, p. 106743. DOI: 10.1016/j.buildenv.2020.106743.
BreathLife (2016): https://breathelife2030.org/solutions/citywide-solutions/ Accessed 2021/03/16
Climate & Clean Air Coalition (2020): World Cities day event focuses on how health, climate and urban air pollution are interlinked, https://breathelife2030.org/news/world-cities-day-event-focuses-health-climate-urban-air-pollution-interlinked/ Accessed 2021/03/16
Das, Ritwajit (2020) How community-based air quality monitoring is helping the city of Bengaluru fight back against air pollution, https://breathelife2030.org/news/community-based-air-quality-monitoring-helping-city-bengaluru-fight-back-air-pollution/ Accessed 2021/03/16
Goodfellow, Ian; Bengio, Yoshua; Courville Aaron. (2016): Deep Learning. MIT Press.
Institute of hygiene and environment, Hamburg. Leuchtbakterientest. Accessed 2021/03/17. https://www.hamburg.de/hu/biotestverfahren/2604448/leuchtbakterientest/
Kang, Gaganjot Kaur; Gao, Jerry Zeyu; Chiao, Sen; Lu, Shengqiang; Xie, Gang (2018): Air Quality Prediction: Big Data and Machine Learning Approaches. In IJESD 9 (1), pp. 8–16. DOI: 10.18178/ijesd.2018.9.1.1066.
Larson, Jeff; Mattu, Surya; Kirchner, Lauren; Angwin, Julia (2016): How We Analyzed the COMPAS Recidivism Algorithm
Liao, Xiong; Tu, Hong; Maddock, Jay E.; Fan, Si; Lan, Guilin; Wu, Yanyan et al. (2015): Residents’ perception of air quality, pollution sources, and air pollution control in Nanchang, China. In Atmospheric Pollution Research 6 (5), pp. 835–841. DOI: 10.5094/APR.2015.092.
Nikolopoulou, Marialena. (2009). PERCEPTION OF AIR POLLUTION AND COMFORT IN THE URBAN, Conference: CISBAT International Scientific Conference, Lausanne
Nisky, Ilana; Hartcher-O’Brien, Jess; Wiertlewski, Michaël; Smeets, Jeroen (2020): Haptics: Science, Technology, Applications. Cham: Springer International Publishing (12272).
Peng, Minggang; Zhang, Hui; Evans, Richard D.; Zhong, Xiaohui; Yang, Kun (2019): Actual Air Pollution, Environmental Transparency, and the Perception of Air Pollution in China. In The Journal of Environment & Development 28 (1), pp. 78–105. DOI: 10.1177/1070496518821713.
Rosenblatt, Frank (1958): The perceptron - a probabilistic model for information storage and organization in the brain.
Smedley, Tim. 2019/11/15. The toxic killers in our air too small to see. Accessed 2021/03/14. https://www.bbc.com/future/article/20191113-the-toxic-killers-in-our-air-too-small-to-see
Sokhanvar, Saeed S. (2013): Tactile sensing and displays. Haptic feedback for minimally invasive surgery and robotics. Chichester, West Sussex, U.K.: John Wiley & Sons.
Spiroska, Jana; Rahman, Asif; Pal, Saptarshi (2011): Air Pollution in Kolkata: An Analysis of Current Status and Interrelation between Different Factors. In South East European University Review 8 (1). DOI: 10.2478/v10306-012-0012-7.
Valueva, M.V.; Nagornov N.N.; Lyakhov, P.A.; Valuev, G.V.; Chervyakov, N.I. (2020) Application of the residue number system to reduce hardware costs of the convolutional neural network implementation, Mathematics and Computers in Simulation. https://doi.org/10.1016/j.matcom.2020.04.031.
VISATON. 2010/04. Basic principles of exciter-technology. Accessed 2021/03/14.
Vu, Tuan V.; Shi, Zongbo; Cheng, Jing; Zhang, Qiang; He, Kebin; Wang, Shuxiao; Harrison, Roy M. (2019): Assessing the impact of clean air action on air quality trends in Beijing using a machine learning technique. In Atmos. Chem. Phys. 19 (17), pp. 11303–11314. DOI: 10.5194/acp-19-11303-2019.
World Health Organization (2006): Air quality guidelines. Global update 2005 : particulate matter, ozone, nitrogen dioxide, and sulfur dioxide. Copenhagen: World Health Organization.
Whitney, Matt; Quin, Hu (2021): How China is tackling air pollution with big data, https://breathelife2030.org/news/china-tackling-air-pollution-big-data/ Accessed 2021/03/16
Wu, Yi-Chen; Shiledar, Ashutosh; Li, Yi-Cheng; Wong, Jeffrey; Feng, Steve; Chen, Xuan et al. (2017): Air quality monitoring using mobile microscopy and machine learning. In Light, science & applications 6 (9), e17046. DOI: 10.1038/lsa.2017.46.
1 Android Bluetooth. Majdi_la. (Sample Code) https://stackoverflow.com/questions/13450406/how-to-receive-serial-data-using-android-bluetooth. CC BY-SA 3.0.
2 Android GPS. Azhar. (Sample Code) https://www.tutorialspoint.com/how-to-get-the-current-gps-location-programmatically-on-android-using-kotlin. Terms apply.
3 Android TFlite. Anupamchugh. (Sample Code) anupamchugh/AndroidTfLiteCameraX. Pending request.
4 FastAPI. Sebastián Ramírez. (Library) tiangolo/fastapi. https://fastapi.tiangolo.com/. MIT-License.
5 I2cdevlib. Jeff Rowberg. (Library) jrowberg/i2cdevlib. MIT-License.
6 Keras: Multiple Inputs and Mixed Data. Adrian Rosebrock. (Sample Code) https://www.pyimagesearch.com/2019/02/04/keras-multiple-inputs-and-mixed-data/
7 Leaflet. Vladimir Agafonkin. (Library) Leaflet/Leaflet. https://leafletjs.com/. BSD-2-Clause.
8 Maperitive. Igor Brejc. (Program) https://maperitive.net. Terms apply.
9 Meteostat. Christian Lamprecht. (DB/Library) https://meteostat.net. CC-BY-NC 4.0/MIT-License.
10 officialAQIus. OpenData Stuttgart. Rewritten by Timo Bilhöfer in Python. (Library) https://github.com/opendata-stuttgart/feinstaub-map-v2/blob/master/src/js/feinstaub-api.js. MIT-License.
11 OpenStreetMap. OpenStreetMap contributors. (DB) https://www.openstreetmap.org/copyright. Terms apply.
12 Overpass-API. Wiktorn. (Docker-Image) wiktorn/Overpass-API. AGPL 3.0.
13 Pandas. Pandas contributors. (Library) https://pandas.pydata.org. BSD-3 Clause
14 Python. Python Software Foundation. (Interpreter) https://python.org. PSF-License
15 sensor.community. (DB) https://archive.sensor.community/. Open Data Commons: Database Contents License (DbCL) v1.0.
16 Sqlite3. (Library & DB Language) https://www.sqlite.org. Public Domain.
17 TensorFlow. TensorFlow Community. (Library & Sample Code) https://www.tensorflow.org. Apache-License 2.0.
18 VisionAir. Harshita Diddee, Divyanshu Sharma, Shivam Grover, Shivani Jindal. (DB) https://vision-air.github.io. MIT-License
This project is imagined and created by Timo Bilhöfer, Markus Pfaff and Maria Rădulescu.
As part of the seminar Learning Atmospheres in the winter-semester 2020/21 it is supported by Irina Auernhammer, Silas Kalmbach and Prof. Lucio Blandini from the Institute for Lightweight Structures and Conceptual Design (ILEK) and is also part of the Collaborative Research Centre 1244 (SFB 1244).