Pedestrian-spotting software for driverless cars improves in speed and accuracy

IEEE Spectrum : To avoid accidents, driverless cars must quickly detect pedestrians and other objects in the roadway. Current systems rely on radar, lidar, and other sensors, but they are expensive and complex. Researchers at the University of California, San Diego, are working on a video-based system that would be cheaper, faster, and more accurate. To analyze the video images, the researchers have combined two computer algorithms. A cascade detection algorithm allows the computer to break an image down into a series of smaller windows and discard those that don’t contain a person or person-like object. A deep-learning algorithm then performs the more difficult task of distinguishing between people and other objects. Although the new algorithm works only in binary mode, the researchers hope to extend it to detect multiple objects simultaneously.