Computer vision classifier and platform for automatic counting: More than cars

Data on urban mobility is traditionally obtained by polling or by people observing and reporting the counting. Even though this task is expensive, labor intensive, and prone to subjectivity, the argument supporting this practice is the high cost of electronic devices such as infrared sensors, loop inductors, or piezo-electric sensors. This project proposes a collective monitoring platform for data collection to overcome the aforementioned issues. Data collection will be through a more sophisticated yet less expensive process. We develop a software that processes existing video streams from security cameras to collect urban mobility data. Our software works not only for cameras from public institutions but also from urban and cycling activists. For privacy reasons, only the transportation's counting total are shared. Camera owners do not need to share video flows into our web server. The web server meets OGC standards for information storage and also allows consultation and public access to gathered data. The existence of geographically distributed and temporally continuous data about the number of cyclists, pedestrians, cars, and buses is expected to reveal the real use of existing infrastructure. The geo-referenced data obtained from our pilot study is available at http://www.tivo.ec:8080/cliente. OpenCV library is used for processing, counting, and generating results in a Raspberry Pi. Results of classification accuracy are not yet available. Nevertheless, accuracy of counting is about 83% and the classification discriminates with success the cars but lack in sorting of other categories yet.