When it comes to the sensor technology for autonomous driving, there always seems two schools of thoughts: beliefs in Lidar/Radar OR beliefs in Vision. I asked two startup founders in autonomous-driving to get their point of views:
- Amit Steinberg, CTO and co-founder of Innoviz, a leading provider of LiDAR remote sensing solutions (Adanim, Israel)
- Dr. Qufei Wang, founder of EVSoft, building sharable driverless cars (Guangzhou, China)
Maggie: what’s the best way to fusing Vision and Lidar?
Amit:There are two main methodologies for fusing inputs from multiple sensors. One is to perform fusion on raw data received directly from the different sensors (LiDARs, cameras, radars, ultrasound). The other is to run higher level computer vision algorithms on the output of each sensor; whereas the fusion is performed on already processed data. We believe that the latter methodology is more robust because each sensor’s raw data processing is independent and decoupled from the other sensors and therefore provides additional backup.
Qufei:I am thinking a very very low-level fusion of Lidar/Radar and Vision–raw data level–where Vision can simplify the design of Lidar and Radar. Taking Radar forexample, the major difficulty of a forward facing Radar is handling the complex response of various objects, such as a small fracture on the road pavement. And Vision is very good at road surface colors. So, instead of using AI to analyze the Radar-generated data alone, I would use data generated by Vision and Radar together. In the same time, one big (maybe the biggest) challenge of Vision is the reflection of glasses, which Radar has no problem with. Therefore, I would only use complex responses of Radar to cross-validate the Vision-generated data. In this way, Radar immediately becomes simple and clear. Vice versa, the difficulty that lies in using Lidar alone is easily solved by fusing with Vision at a low level. Fusing at raw data level is a very exciting topic. It is also suitable for the current AI technologysuch as Convolutional Neural Network.
Maggie: is it possible to reduce the cost of Lidar down to around $200 by reducing the requirement such as making low line count and low accuracy Lidar?
Amit: In theory, one can reduce spec and thus reduce cost. Nevertheless, such LiDAR will not be able to meet any of the automotive market requirement and be considered as an automotive grade. Thus — such LiDAR will not enable higher levels of autonomy and will not advance the market towards enabling full autonomy. Innoviz is providing a comprehensive mass-market solution and currently is the only LiDAR provider to deliver a solution that enables high performance, safety, reliability at a lower cost and smaller footprint. In addition, our technology has won validation from leading stakeholders in the automotive industry, including BMW with the recent announcement of their selection of our LiDAR technology for series production of Autonomous Vehicles in 2021. Early samples of our auto-grade InnovizOne LiDAR will be available in 2019. Initially priced at a few hundred dollars, Innoviz plans to drive these prices down further as volumes increase.
Qufei: With low line count and low-resolution Lidar may be able to reduce the Lidar price. If fused with Vision and Radar, the lower cost Lidar, as a system, can still have reliable quality. Car makers, at least those in China, are open to new designs, especially to those with much lower prices. Car makers judge the quality and reliability as a whole, not just by the Lidar component alone. For quite a few years, the price of Lidar remains forbiddingly high. It is time to take some fundamental changes. Without such fundamental design scheme shift, Lidar is unlikely able to reduce cost, therefore, fusing with vision at a very low signal level has a chance to reduce total costs.
P.S. A few links for references
Lidar Just Got Way Better — But It’s Still Too Expensive for Your Car (November 28, 2017, MIT Technology Review)
Why doesn’t Tesla use LIDAR?(January 6, 2017, tesla.com)
Lidar startup says it can mass-produce cost-effective sensors(April 12, 2018, Automotive News)