1
$\begingroup$

I have a liberal arts background so I need help understanding this paper, particularly pages 26 to 30. The authors test a four-camera system for localization, mapping, and obstacle detection for self-driving cars. The paper seems to say the multi-camera system can map the environment to within an average of 7 cm (2.8 inches) of accuracy (with the largest error being 16 cm or 6.3) and detect obstacles to within 10 cm (3.9 inches) of accuracy. Am I getting this right?

Given that automotive lidar can detect objects to within 1.5 cm (0.6 inches) of accuracy, and given that for driving purposes the difference between 1.5 cm and 7 cm, 10 cm, or 16 cm seems quite small, can a multi-camera system be used instead of lidar in a self-driving car application? How do driving speeds affect things? What crucial elements of the problem space might I be overlooking or misunderstanding?

$\endgroup$
2
$\begingroup$

It seems that LIDAR presents a problem for resolving the car's environment at higher speed. While I'm not too familiar with the dynamics of LIDAR I do know that it's a physical system that relies on sending and receiving laser pulses to various points around the car by way of rotating mirrors. As speeds increase, it seems different arrangements of mirrors and light collectors might have to be used to maintain a high-resolution image. There's some evidence that Doppler LIDAR (developed in the 1990s) became less accurate with higher velocities. However, LIDAR is partly preferred over radar because of its higher accuracy even when tracking objects at high speeds - this is why LIDAR guns are increasingly being used by police instead of radar guns to track speeding vehicles. It seems natural that a set of high-resolution cameras paired with a well-trained neural network would not be subject to the same physical limitations as LIDAR.

I think that an important intuition to consider is that while LIDAR is used to generate clouds of data points whose shapes and patterns can be analyzed by autonomous car software, cameras can pick up non-topograhical features such as road lines, the content of roadsigns, and additional location context such as storefronts and intersection layouts. Considering that these cameras can use pattern recognition and stereoscopy to also generate a 3D topographic map of the environment, it seems plausible that Level 5 self driving cars would not require LIDAR.

Here's an interesting look at the problem.

$\endgroup$
1
$\begingroup$

Given that a self-driving car is trying to replicate the performance of a two-camera system (or one in a pinch), there is nothing in principal that mandates lidar for a self-driving car. Lidar is a shortcut, substituting sensor sophistication for image-processing sophistication. AFAIK Nvidia's own self-driving vehicle doesn't have Lidar. My personal opinion is that Level 5 self-driving vehicles won't be practical until they have the kind of image-processing sophistication that makes Lidar an unnecessary crutch.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.