Why visual SLAM is seldomly used in autonomous driving?

Question

In research called Occupancy Network by Tesla, the occupancy map is constructed by the deep learning method.

We know that visual SLAM can compose maps like Octomap, either. And Octomap can be used for quadrotor's fast flight (can refer to this RAPTOR article, a TRO).

However, we can seldom see visual SLAM projects that have been used in the autonomous driving area (or maybe because I don't know, if you know any, please tell me, big thanks!!! ).

What's the reason for it? Because visual SLAM can only construct sparse point clouds? And the time cost for constructing dense point clouds is heavy? Or are there any other reasons?

Answer 1

There are a couple of reasons.

1. Autonomous driving is all about perception and localization. Not the mapping. For localization, a simple map and GPS are enough. In terms of perception, the conventional SLAM is not intended for it.

2. Road is changing every day. The conventional SLAM usually can't localize itself when the scene has significantly changed. Also, companies don't want to update their point cloud map every day.

There are another area where visual SLAM is used for autonomy such as logistics robot, robotic lawn mower, cleaning robot where small scale indoor localization using camera is just enough for their tasks.