Computer Vision for monitoring the condition of streets and the city environment
Use data from cameras on cars controlling payments in the streets to detect garbage in the streets, potholes and other anomalies which should be removed.
When you need to sort waste, your place for sorting is outside your household. You need an overview of the sorting place, which could be overfilled. So when you take your paper, plastic bottles or glasses and the containers are full, you must take them back or lay them near them, which looks messy and demotivates you from sorting waste. Computer vision model for detecting overflowing containers These detections could be used as information for the people which place for waste separation they can use(is not overflowing). And also for City Service to manage sooner sorted waste disposal.
We used ResNet-32 and Finetune learning. The solution could be extended to other use cases, such as monitoring potholes, broken signs or other street anomalies. Since we didn't have access to the videos from cars, we made Acc 200 images of containers with phones in the streets, and another 100 hundred was obtained from the internet. We prepared a simple frontend application in React using a leaflet framework for monitoring results on the map of the City.
A production-ready solution with more than 95% accuracy needs 60MDs of annotation work and 15MDs for training finetuning. 40MDs for proper working frontend and backend for visualisation. 20MDs for data engineering for pipeline processing data. 30MDs for integration with City systems. The solution was created in cooperation with colleagues from the Datasentics company.
The project is in the process of implementation yet.
Dependencies and constraints
The only constraint is access to video data from car cameras that monitor parking payments.