The Celestini Program’s objective is to grow the next generation of networking and communications innovators in developing countries by empowering undergraduate students to create social and economic transformation in their communities through technology. The Marconi Society and our Young Scholars achieve this by selecting universities with promising telecommunications and engineering undergrads and providing them with support and mentorship to help tap their students’ true potential.
Two Years, Two Big Issues
In 2016, the Celestini Program started in Uganda and in 2017 it expanded to India under the leadership of young scholar Dr. Aakanksha Chowdhery (Google) in partnership with Prof. Brejesh Lall (IIT Delhi). Student teams are selected through a technical screening exam, as well as an interview in spring. The selected student teams work over the summer at IIT Delhi to identify an important problem in their community that they would like to solve and then to prototype their projects. In 2017, the first year of the project pilot, thirty students expressed interest in working on the project and six students were selected to work on the problem of improving road safety.
In the second year of the project, 2018, more than one hundred students expressed interest in working on the program during the summer and three student teams, comprising eight students, were selected. They chose problem statements related to air pollution and road safety in Delhi. One team prototyped a website that forecasts air pollution levels in Delhi over the next 24 hours. Another team designed an Android app that uses a user-uploaded photos to estimate air pollution levels. A third team prototyped a low-latency platform to transmit vehicle-to-vehicle alerts about potential road safety hazards/collisions using Xbee radios.
Unique Approaches to Improving Air Quality
Preserving air quality is a critical challenge in the industrial and urban areas of many emerging economies. According to the World Health Organization (WHO) Global Ambient Air Quality Database released in Geneva, India has 14 out of the 15 most polluted cities in the world in terms of PM (particulate matter) 2.5 concentrations. One potential solution is to increase awareness of the problem by enabling users to understand and track the level of air pollution, as well as to receive forecasts for the next day along with information about the potential sources of pollution. This could provide the basis for taking effective pollution control measures.
One of the student teams worked on designing a temporal forecasting solution to predict the real-time and fine-grained air quality information in five locations around Delhi, based on the historical data reported by Central Pollution Control board. The solution predicts the air quality over the next 24 hours based on the level of different air pollutants including sulphur dioxide (SO2), nitrogen dioxide (NOx), PM2.5 and PM10. It also tracks the seasonal variations of the major pollutants and the potential sources of pollution at different points in time. The challenge in their project is to build accurate machine learning models that effectively adapt to the daily and seasonal variations of various pollutants. They developed an advanced machine learning model called CLair using LSTM techniques. The team deployed their solution approach on the Google Cloud Platform to automatically generate predictions every few hours on this website for five Delhi locations.
They showcase their project in this video.
Temporal forecasting can predict air pollution levels in locations where the Delhi government provides air quality data hourly, but this is limited to specific locations. A scalable approach requires crowdsourcing where we use inputs from the entire population and the easiest way to leverage crowdsourcing is via smartphone applications that are widely used by Delhi residents. Toward this goal, one of the student teams developed an Android smartphone application called Air Cognizer.
This application allows users to upload an input image of the sky horizon taken from their smartphone camera. Based on the certain features of the sky, such as how blue it is, the app predicts air quality particulate matter indicator, PM2.5 concentration, with an error less than 5%. The application combines image processing with machine learning using Tensorflow Lite to generate estimates by combining a pre-training machine learning model with a model trained online for each location based on all the user-uploaded photos. Two key challenges that students solve are preprocessing the data collected from different smartphone cameras so that the machine learning model works accurately and deploying this machine learning model on the smartphone with Tensorflow Lite to enable a low-latency real-time prediction experience. The video below provides a sample of the Android application that the team has launched in Google Play store
Year Two of Improving Road Safety
Another challenging problem that the student teams worked on is road safety. Over 200,000 people in India lost their lives in road accidents in 2015. Traffic accidents are the top cause of death for people aged 19-25. In Delhi, many of these accidents are caused by buses hitting pedestrians and cyclists. In 2017, one student team worked on prototyping a solution using Raspberry Pis and dashboard cameras to detect pedestrians and cyclists on potential collision course with the vehicle showcased in this video.
This year, a student team prototyped a solution that allows multiple vehicles to talk to one another at low latencies (tens of milliseconds) to send real time alerts about possible impending collisions to drivers behind them to prevent chain-reaction car accidents. This system leverages computer vision to classify a given scenario as one that may result in collision. Then the system uses vehicle to vehicle communications to broadcast alerts. Each vehicle, acting as a node, broadcasts information about its speed, location etc. and the other nodes receive and process this information based on the degree of relevance that the message holds. The solution approach was designed over Xbee radios as a low-latency solution (~30-40ms) available off-the shelf at low cost. This video and website link showcases their work during the project.
The concluding ceremony of this year was held on November 1 at IIT Delhi where Marconi Society board member, Prof. Andrea Goldsmith (Stephen Harris Professor of Electrical Engineering, Stanford University), gave an inaugural address. The ceremony was attended by the IIT Delhi Dean of Alumni Affairs & International Programmes, faculty members from EE and CS and industry partners. The winning team who developed the smartphone application Air Cognizer for air pollution analytics using smartphone camera photos was awarded a cash prize of $1500.
Our results so far show that the Celestini Program brings hands-on learning and critical decision-making skills to participating students and that we are inspiring these students to pursue STEM-related studies to solve the problems they see in their communities. The participating students value the experience of developing prototypes that solve real-world problems with the potential to improve the well-being of their communities. We have already had the privilege of working with 14 outstanding students in the Celestini Program India and look forward to seeing the exceptional ways that they use STEM in their careers.