Computer Vision News - September 2018

Computer Vision News Application 12 Application: ZOOX Zoox describes itself as a group of “ inventors, builders and doers ” and, as a company that develops fully autonomous vehicles from scratch, they surely live up to the description. The startup has already raised nearly $800 million in funding. Their end goal? To provide self-driving vehicles as a service, i.e. robot taxis. While certainly not the only company working towards autonomous driving, Zoox stands out compared to others because, rather than building on existing vehicles, they have actually redesigned the entire idea of transportation by building a robotic vehicle from the ground up. The idea started with Zoox founders, Tim Kentley-Klay and Dr. Jesse Levinson . Kentley-Klay, a designer originally from Australia, considered the tremendous potential of this type of technology. He reached out to Levinson, who at that time was leading the Stanford Autonomous Driving Project . They met, and together, eventually began collaborating until Zoox was finally born. The Lidar team manager David Pfeiffer , along with the company’s Director of Detection and Tracking, Sarah Tariq , share fascinating insights with us into this new autonomous vehicle and the future of self-driving cars. David says that technology completely changes the setup of a car as we know it today. Their system allows them to get rid of all extraneous parts needed in a car driven by humans which are now obsolete in a fully autonomous configuration. For example, their model allows passengers to sit facing each other, rather than all facing forward. The perception technology fuses three types of modalities: radar, lidar, and cameras . David explains why the cars require this combination: “ Why all three? Every sensor has different strengths and weaknesses. The camera has magnificent resolution and is really good at classifying objects, but lacks depth. Lidar has excellent depth but is somewhat sparse, and tracking is somewhat complex. Radar doesn’t have great lateral resolution, and only limited accuracy in terms of saying where objects are. But it’s very good with velocity and works equally well in all weather conditions. All of them are, in their way, perpendicular to another. You want to guarantee maximum availability of objects when you detect them. It’s a bit of a change in paradigm. Let's consider a traditional off-the-shelf collision warning system: even when the car beeps for no reason, the driver is always in charge to decide what to do, right? David Pfeiffer