A world where everything is automated should be perfectly safe, in theory. In this reality, vehicles travel along predisposed channels and factories run themselves, bringing the potential for collisions, congestion, and other incidents very close to zero. We don’t live in that world, however, and so long as people are around, robots will have to act accordingly.
A group of researchers from the University of California, San Diego (UCSD) have devised a new way for machines and humans to coexist, and it comes in the form of a pedestrian detection system that is much more precise than those being used today. Developed at the university’s Statistical Visual Computing Lab, the technology has a variety of applications, the most notable of which is helping self-driving cars see more clearly.
While conventional pedestrian detection systems scan everything in the vehicle’s path for people, the UCSD team’s design leans on cascade detection, which automatically cuts out areas that it can easily recognize as empty — the sky, for example. Then, a unique algorithm scans person-like shapes such as trees, using deep learning and complex pattern recognition to eliminate false alerts.
Because the system only has to focus on small detection windows, it can spot people at a rate of 2 to 4 frames per second (near real-time), resulting in much greater accuracy and response time overall. According to the engineers behind the project, their versions make half as many mistakes as are made on the road right now.
As of now, the system can only recognize one type of object at a time, so there’s a long way to go before it’s ready for production. That being said, the groundwork has been laid, and eventually, the UCSD team could revolutionize pedestrian detection in cars.
