Apple self-driving cars are back! VoxelNet may drive the autonomous vehicles

The cat is out of the bag. There is no secret in why Apple kept denying its interest in building self-driving cars – it wanted to keep it a secret. Just like the way it held a behind the scenes discussions with Tesla founder Elon Musk three years back.

For all the hype around, Apple does have a permit from Californian authorities to test self-driving cars. And Tim Cook has been on record calling self-driving cars "the mother of all AI projects.”

Whatever may have happened with the Project Titan, Apple seems to have shifted gears on the autonomous vehicles proposal, focusing more on the software side of the equation. Now the company is working on using a light-based technology to make it easier for self-driving cars to identify pedestrians and cyclists.

In a new research paper published in academic repository Arxiv, Apple computer scientists Yin Zhou and Oncel Tuzel have discussed a new object detection method for self-driving systems based on LiDAR (Light Detection and Ranging) – a method to gauge distance by illuminating a target with a pulsed laser light, and measuring how long it takes to return.

The research paper describes a method for using machine learning to translate the raw point cloud data gathered by LiDAR arrays into results that include detection of 3D objects, including bicycles and pedestrians, with no additional sensor data required.

This new way to use LiDAR is what the Apple researchers call VoxelNet: an end-to-end trainable deep architecture for point cloud based 3D detection. A voxel is a point on a 3D grid.

Accurate detection of objects in 3D point clouds has been a central problem in many applications. And most existing methods in LiDAR-based 3D detection rely on hand-crafted feature representations, for example, a bird’s eye view projection. But with VoxelNet, this manual feature bottleneck is removed.

So how does VoxelNet detect small obstacles using the LiDAR sensing method? The researcher duo elaborated that Voxelnet “divides a point cloud into equally spaced 3D voxels and transforms a group of points within each voxel into a unified feature representation through the newly introduced voxel feature encoding (VFE) layer.” In this way, the point cloud is encoded as a descriptive volumetric representation, which is then connected to a RPN (region proposal network) to generate detections.

But LiDAR comes with its own limitations. While it is great at figuring out the exact position of objects in 3D space, it does have a notoriously low resolution. Which is why firms have been using different methods to overcome LiDAR’s shortcomings – some even using regular cameras for object identification. Tesla does not even use LiDAR at all.

In order to figure out exactly what the object is, vehicles must therefore also rely on other sensors and cameras. And that adds costs and processing bottlenecks. Apple’s research is still in early stages, and the company is contemplating putting its own software suite at the end of the LIDAR sensor itself (which it claims greatly increases its effectiveness). The ‘complete picture’ will hopefully be out in days to come.

While the scientific part of the paper is interesting, the fact that Apple has come out with its first external publication on driverless vehicle projects has finally put things out of the closet. Arxiv is often used by researchers to get preliminary feedback before publishing in a final form, and Apple’s paper proves it can no longer go solo on the challenging research subject of AI and self-driving. The tech pioneer doesn’t want to take chances anymore after the earlier fiasco.

After all, Apple is self-admittedly working on its “next big thing.” That could potentially disrupt the automobile industry like never before!