Researchers in the Department of Computer Science and Engineering (CS&E) are continuing their work on a vision-based system for counting pedestrians and bicyclists. The system uses machine learning principles and complex algorithms to process video data and classify objects in the scene as either cyclists or pedestrians.
This vision-based system overcomes many of the shortcomings of existing detection and counting technologies—such as loop counters, buried pressure pads, and infrared counters—because it is capable of distinguishing between a cyclist and a pedestrian. This ability allows the system to obtain accurate traffic counts on bicycle trails, bridges, bicycle lanes, and other locations with heavy bicycle and pedestrian traffic.
The system works by first filtering out the background of available video data. It then takes the remaining foreground objects, called blobs, and converts them into small image patches. These patches are compared to object dictionaries for people and bicycles that the system has previously “learned” in a training mode. It uses these dictionaries to determine the classification for each image patch, and then to make an overall classification of the entire blob as a bicyclist or pedestrian.
The research team tested the system by acquiring two hours of high-quality video data from three walkway sites at the University of Minnesota, as well as some high-velocity bicycle traffic data from the Gateway State Trail in Uptown Minneapolis. Results indicate that the system was able to correctly identify bicyclists 86 percent of the time and pedestrians 98 percent of the time. The overall accuracy of the system was 96 percent.
Download the full study here.
Other than the obvious blunder that the Gateway State Trail is not anywhere near Uptown (one would assume they mean the Midtown Greenway), it sounds like a cool project.
Hopefully this, or some other, method of counting can become reliable and cheap enough that we can begin collecting continuous data. We can already do this somewhat using loop detectors or infrared detectors. However, loop detectors don’t count pedestrians, and (if memory serves) infrared detectors don’t do well distinguishing between bikes & peds.
What other automated counting strategies are there?