RESEARCH: I’m walking here!
In order for a self-driving car to safely avoid pedestrians, it needs to be able to predict where they will head—whether they will walk directly in front of its path or stop to let it pass by. This problem, known in academic speak as “pose and gait estimation,” is a particularly hot topic in the research community, and many solutions have been tested over the years to produce a model with the highest possible accuracy.
Now researchers at the University of Michigan have the latest proposal. In a new paper published in the IEEE Robotics and Automation Letters, they demonstrated a new deep-learning technique that is trained on lots of video of pedestrians but also given an understanding of the biomechanics of how we walk. The resultant model can then produce a full-body, three-dimensional rendering of where pedestrians will move, and works even when there’s more than one person in the frame.
When the researchers tested the model against a standard dataset of pedestrians crossing a real urban intersection, it outperformed previous methods. The Algorithm from MIT Tech Review firstname.lastname@example.org