apollo self driving car

apollo self-driving car

HD maps

where to define the center-meter fidelity level maps, with landmarks, and even the height dimension as well, which is used as the guiding map and also the global coordinate to locate the vehicle in the world.

how to keep HD maps updated is a big invest. Baidu has hundreds service cars in China to collect the natural highway HD map data, even still to 2020.

Localization

SLAM tech is a robot maping the world at the same time localizing itself in the world. with HD maps prepared first, self-driving car only need localize itself at every timestep. and the common idea is comparing a few local landmarks with the corresponding global landmarks, then transfer the local vehicle position to its global position in HD map.

LiDAR localization

either based on the cloud points matching from continous timestamps; or calculating the error between the LiDAR points with the HD maps points; or based on the Karman filter, which give the highest possibility of the location of vehicle

always accessible but not easy to construct, especially requiring HD maps.

Visual localization

based on particle filters, which give the most likely location of the vehicle.

Perception

detection && classification

where the object located, and classify it. detection CNN to find the object in the image; then use classification CNN to classify it. or use a combined CNN to detect and classify at same time.

tracking

tracking helps when detection failed; also tracking helps to identity the object, so when objects overlapped in the image, still can tell which is which.

Prediction

model based prediction

data driven prediction

after prediction, then generate the trajectory.

Planning

transfer the world to graph, and find routings in the graph. routing is a global path, comparing to trajectory is the local way.

Vehicle control

linear quadratic regulator

PID

model predictive control