The idea of cars that can drive themselves has been floating in and out of the dreams of humanity’s great thinkers for decades. But with recent advances in the fields of computer science, applied mathematics, and deep learning, this idea is looking less like a dream and more like a reality within our reach.
Most current attempts at autonomous driving rely on some advanced hardware to create a constantly updated map of a vehicle’s surroundings. Using a technology called LiDAR – in essence a kind of radar using lasers – an autonomous driving system can “paint” its surroundings to generate 3D models of objects around it.
A San Diego-based startup, TuSimple, was founded on a different approach.
“What if instead of using LiDAR arrays which cost hundreds of thousands of dollars, we could use simple cameras to see the road like humans do?” co-founder and chief technology officer Dr. Xiaodi Hou explains.
Dr. Hou is a world leader in the field of computer vision. With a PhD from Caltech, he gathered some leading scientists in computer disciplines from across the world, brought them to San Diego, and sat down to design an artificial intelligence system that can “see” like we do.
“The problem is that computer vision is orders of magnitude more complex,” Dr. Hou says. “It needs to do things which come effortlessly to humans – identify objects, predict behavior, discriminate between significant and insignificant, vehicle and pedestrian, road and environment.”
But once a computer vision-based AI is created, the hardware needed to install it in a vehicle is much cheaper than a LiDAR approach.
Of course, making a computer program that can interpret visual information like a human is easier said than done.
“We are using computer vision as the basis for our autonomous driving platform, but this is the forefront of innovation in this area. The scientific community hasn’t caught up yet – the literature isn’t there, the references aren’t there, the problems haven’t been defined. We are essentially creating this field each day from scratch,” Dr. Hou explains.
TuSimple first broke the news last fall, when they won ten first-place rankings in the computer vision benchmark dataset competitions released by KITTI and Cityscapes. Now they are offering their own computer vision benchmark dataset – the world’s first dataset specifically oriented toward automated driving.
“Benchmark datasets like KITTI, Cityscapes, and now ours foster innovation and creative solutions. We believe that development in the field of AI is a collective process, not a zero-sum game, and any growth in knowledge or processes improves the entire field of study,” Dr. Hou adds.
With cash prizes of
$1000 for first place,
$500 for second place, and
$250 for third place in the competition, they hope to foster innovation, to encourage creativity, to inspire novel solutions, and to bring the world’s community of computer vision researchers together to advance the whole field. They will also formally recognize the winning teams at this summer’s upcoming CVPR 2017 Conference in Honolulu, HI from July 21-26, which TuSimple is co-sponsoring.
Instructions and all data can be found on their website