Monitoring Cleanliness of Public Transportation with Computer Vision

Due to the ongoing global trend of urbanization, the demand for public transportation is ever-increasing. In order to answer to the growing demand, public transportation vehicles will likely be autonomously operated in the near future, creating a need for advanced surveillance systems. In this study, a computer vision system capable of evaluating the cleanliness of a public transportation vehicle interior is presented. In a laboratory setting mimicking a real vehicle interior, the presented system was studied with a traditional ceiling-mounted wide-angle camera, as well as a linearly actuated camera providing wider visual coverage of the environment. Image data was gathered of the setting in a clean condition as well as containing common trash objects. With this data, the evaluation accuracy of the system was measured with background subtraction, edge detection with random forest, SURF with k-means clustering and XGBoost, and Single-Shot multibox Detector. Single-Shot multibox Detector was found to be the most accurate approach, reaching 95-100% accuracy reliably. The linearly actuated camera was not found to provide notable difference in accuracy compared to the wide-angle camera. Overall, the results indicate that the presented system could be applied in cleanliness monitoring.