Computer vision in the service of SmartCity of Brno

On 9 April, a team of enthusiasts from DataSentics based in the South-Moravian region took part in the hackathon #hackujbrno organized by the City of Brno, Czechitas and Brno.AI. As the name of the hackathon suggests, the goal was to come up with ideas on how to improve the lives of Brno citizens. For this purpose, the teams could use the newly published open data processed and made available by the data analytics team of the City of Brno. The criteria were not so much the technical skill of the competitors but rather feasibility, innovation and benefit to the city and its inhabitants.

The winning team came up with many ideas for improving Brno city. From monitoring public transport load to detecting stowaways to optimizing traffic lights. In the end, however, they decided to bet on AI expertise developed at DataSentics. Expressly, they set out to capitalize on footage from cars checking parking. The use of Machine Learning and Computer Vision for license plate reading is relatively common, so the team decided to take it to the next level and use a similar approach to solve other challenges, such as detecting potholes, broken road signs, and street width, or overflowing trash cans? They focused primarily on seeing overflowing bins for sorted waste. This knowledge of the current situation could help the local collection and processing of waste in Brno in the future and the citizens who would not waste their precious time looking for free containers. They could find out the current situation, for example, by extending the existing Brno Citizens app. 

Every machine learning problem has three essential ingredients: a well-defined task, data and a model. As our solution’s first ingredient (a well-defined task), the team decided whether the photo shows a full or empty waste collection container, i.e., binary classification of images. Obtaining the second ingredient (data) was a bit more laborious, as guys could not get records directly from the monitoring vehicle during the hackathon. So they acquired the data directly in the field. To simulate the real ones from the video footage, they tried to keep the same height, distance and angle as the monitoring vehicle. The result was about 200 images of different containers; after discarding the unsuitable ones, they were left with 132 images of empty containers and 24 photos of overcrowded containers. They supplemented the overcrowded dataset with 50 images obtained from the Internet. To get the third ingredient (the model), the team used Azure Computer Vision. Considering the training data, they received great results with 81% accuracy. This means four out of five processed images would be determined correctly. This result is easy to improve if enough data is available.  

To give the participants a closer look at the team’s vision, i.e. how to use this real-time information, they created a prototype web application during the hackathon showing the position of the separate waste containers, including their current status (ok/overflowing). For this purpose, they could use Brno’s open data, specifically the containers’ dataset for separated waste.

At the hackathon, the team showed that their project’s data could help solve the overcrowding of containers for separated waste. Luckily, their vision does not end there. The same images can be used to detect, for example, defective road conditions or road signs. This could bring Brno closer to cities at the global cutting edge of Smart City solutions, such as Amsterdam or Las Vegas, where they similarly use imagery of volunteers or supplies. So, given the great results from the hackathon, the overall first place and the special mention in the AI category.