Sponsor: Department of Transportation, Great Streets Program
University: California State University, Los Angeles
Status: In progress
The City currently does bicycle and pedestrian counts via having a person manually count the number of cyclists and pedestrians that go through an intersection via a video capture. However, thanks to advances in machine vision technology, we can now automate that, allowing us to constantly count the number of pedestrians and cyclists, rather than only doing a 24-hour sample. This project is a proof of concept of how we could count the number folks moving through our streets. CSULA students are working on the project.
Other video recognition software for counting active transportation activity has proven ineffective due to either technological or cost limitations for the City of LA. Most of these algorithms are trained on datasets that don’t look like LA (for example, we don’t get snow here). When we tested them, the proved ineffective. This project extends on work for ATSAC and RIITS regional transportation planning software and hardware.
Our solution represents an innovative use of existing resources, local expertise and partnerships to deliver a truly best of breed solution to LADOT. New cameras and equipment to use existing video recognition software systems would be cost prohibitive to install, would require commitment to long term funding of the particular selected vendor, and would lack to ability to “train” an algorithm on specific LA data (rather, we would be limited to the built-in algorithm of the system, which may not have been developed in a similar environment to Los Angeles streets) We tested the algorithm developed in the City’s Pedestrian and Bicyclist Recognition Project against off-the-shelf computer vision technologies provided by Cloud vendors such as Google and Microsoft and found that the off-the-shelf technology failed to accurately recognize pedestrians.
Finally, the paper that was written based on this research was present at the International Conference on Signal and Image Processing-UGC Listed Proceedings (SIGNAL 2018) conference in Paris this year as a “novel application of machine vision in the real world”.