19 Jul Crowdsourcing Machine Learning Data Collection for Traffic Safety
The amount of deaths caused by traffic accidents each year is staggering. In 2014 alone, there were over 35,000 motor vehicle related deaths in just the United States and it’s nearly forty times that worldwide.
With recent advancements in technology, Victor Bahl, director of Microsoft’s Mobility and Networking Research, decided to take this issue head on and use machine learning to make intersections safer. GeekWire recently reported on the ‘Vision Zero Partnership’ composed of Microsoft, the City of Bellevue, the University of Washington and the Washington Department of Transportation, along with other organizations, to achieve the enormous goal of completely eliminating traffic deaths and injuries.
Even with so many organizations involved with this project, they still need everyone’s help to reach this goal. There are tens of thousands of traffic video files to review, so they are turning to crowdsourcing for the review process. Volunteers are helping identify objects in video files in intersections around Bellevue to help Microsoft’s algorithm learn the difference between a bicycle, a car, and a pedestrian, while also being able to recognize near-misses and accidents. By this process of selective learning, they will be able to train the computer to predict areas of danger within intersections, allowing officials to make changes in the intersection before anyone is injured. In effect, they will be avoiding traffic accidents from ever occurring.
Identifying objects in video files is a simple process anyone can do.
I decided to try this out for myself here in the office. There was no sign up or red tape to go through, I was immediately served a video file a few seconds long of an intersection in Bellevue on an ordinary rainy day.
I began by identifying a pedestrian by drawing the box around the person. I checked the “Pedestrian” box on the left to identify the “Object” and then checked the “Crossing the Road” box to indicate to the computer the actions this person was taking in the video file.
I did the same for cars and buses, making sure the boxes drawn around the objects were as tight and accurate as possible. This video file had around twenty objects to track and it took about twenty minutes to complete.
Once I had all of the objects identified in the video, I submitted my work. My lunch was extra satisfying knowing that I had contributed a little bit to help this project reach its ultimate goal. Plus, it was neat to experience another form of selective learning in the exciting field of machine learning and predictive analytics.
The process of identifying the objects in the video was very simple and straightforward which made the entire experience much more enjoyable. If you’re interested in volunteering your time to help identify objects in these videos, click here and you’ll be served a video immediately.
It’s easy and it’s fun to experience the backend of this machine learning project plus you’ll be giving your time for an important cause and exciting project.