MIT grad swaps lasers for cameras, enhances car vision

In its own tests, its proprietary camera system outperformed LiDARs in multiple conditions.
Ameya Paleja
Stock image depicting the use of cameras for computer vision in an autonomous car
Stock image depicting the use of cameras for computer vision in an autonomous car


After getting his Ph.D. from the Massachusetts Institute of Technology (MIT), Leaf Jiang spent more than a decade building laser ranging systems for the military for various 3D sensing applications. In his experience, Leaf found that laser-based detection systems were too expensive to be deployed on autonomous vehicles being developed for the future, and that's how NoDar was born.

Light detection and ranging (LiDAR) systems use laser beams to scan their surroundings and create 3D images from the data obtained when surfaces reflect the light. As companies look to make autonomous driving more mainstream, they rely heavily on LiDAR systems for imaging roads and helping cars make critical decisions on whether an object is a branch of a tree or a human being.

Each LiDAR system can not only cost tens of thousands of dollars but is also not wholly accurate and prone to failure. Leaf's startup, NoDar, promises a much cheaper alternative based on one of the most ubiquitous electronic devices, the digital camera.

How can a camera replace LiDAR?

Camera-based 3D vision has been attempted multiple times before and failed miserably. Unlike LiDAR, a camera-based system depends on ambient light to create images. The results of such imaging vary depending on the time of the day and are often poor in low-light conditions such as rainy or foggy weather.

NoDar, however, claims that camera technology has improved over the years. It has also developed proprietary software to get better results from cameras and ensure that they beat LiDAR systems hands down.

NoDar uses two cameras placed well apart on a vehicle to gather separate views of the road ahead. The two views also allow it to triangulate the location of an object in view and determine its distance from the vehicle. This approach has been used earlier and required precise calibration to be accurate.

Leaf's startup offers a software solution by auto-calibrating the cameras and syncing up their frames. The company has patented the technology for this calibration, and the algorithm can be run in real-time on the chips that cars are equipped with, so no additional hardware is required.

How well does it perform?

To determine the performance of its technology, the startup conducted tests at a remote airstrip in Maine, away from light pollution. Two 5.4-megapixel cameras were placed nearly four feet (1.2 m) apart, and the imagery obtained was compared with a high-end LiDAR system.

NoDar found that its system generated 40 million data points per second in broad daylight against 600,000 LiDAR, an IEEE Spectrum report said. The team also worked with an automobile simulation chamber that could recreate conditions like rain and fog for these tests.

Under conditions of extremely heavy rain, the number of data points dropped by 30 percent, but that for the LiDAR system was 60 percent. In limiting conditions of fog where visibility was only about 145 feet (45 m), the camera-based system still managed accurate measurements for 70 percent of the distance. In contrast, LiDAR-based measurements were accurate for only 20 percent.

In night conditions, which is the true test for the system, NoDar still outperformed LiDAR systems by being able to spot a nearly five-inch (12 cm) piece of lumber from over 400 feet (130 m) away. The high-end LiDAR could spot it only when it was 164 feet (50 m) away.

Leaf is confident of delivering a NoDar system at a fraction of the cost of a LiDAR system. Experts, however, pointed out that a LiDAR system provides a 360-degree view of the car, while NoDAR works only in one direction. Obtaining a similar view would probably require more cameras and computation, which could effectively increase the cost of such a system.

Last month, Interesting Engineering had also reported researchers were using heat signatures for a similar application.

Add Interesting Engineering to your Google News feed.
Add Interesting Engineering to your Google News feed.
message circleSHOW COMMENT (1)chevron
Job Board