Skip to content

Latest commit

 

History

History
189 lines (175 loc) · 7.88 KB

Apollo Self Driving Car.md

File metadata and controls

189 lines (175 loc) · 7.88 KB

Notes From Udacity Free Course

Mapping

  • HD map: high definition map, cm level precision.
  • ROI: region of interest.
  • GNSS: Global Navigation Satellite System, GPS is the most popular GNSS.
  • GPS: frequency is about 10 Hz, 1-3 meters precision.
  • RTK: Real time kinematics positioning.
  • IMU: Inertial Measurement Unit.
    • is used by survey vehicles to map the environment.
    • Accelerometer + gyroscope.
    • high frequency update, up to 1000 Hz.
    • Motion error increases with time. Need to be combined with GPS.

Localization

  • LiDAR Localization

    • ICP (Iterative Closest Point)
      • For a point in 1 scan, find another matching point in the other scan.
      • Compute average distance error by rotating and translating the point clouds.
    • Histogram Filter
      • Also called Sum of Squared Difference (SSD).
      • Slide the point cloud scan from the sensor across every position on the map.
    • Kalman Filter
      • Predict our current state based on our past state and new sensor measurements.
      • Predict State <-----> Update Measurement

  • Visual Localization

    • Uses Particle Filter.
    • Images are easy to obtain.
    • Lack of 3D information and need to rely on 3D maps.

  • Apollo Localization

    Kalman Filter                       Inertial Navigation  
                          prediction  |         ^  
                                      V         | update
GNSS Localization         ->        Position, Velocity  
LiDAR Localization        ->        Position, Heading
    • Inertial Navigation solution is used for the prediction step of the Kalman Filter.
    • GNSS and LiDAR are used for the update step of the Kalman Filter.

Perception

  • Perception task for a self-driving car:
    • Detection: where an object is in the environment.
    • Classification: what the object is.
    • Tracking: observing moving objects across time.
      • tracking handles occlusion (fail of detection), preserves identity.
    • Semantic Segmentation: match each pixel with a semantic category.
      • FCN: fully convolutional network
        • Replace the flat layers at the end of a traditional CNN with convolutional layers
        • Encoder + Decoder
  • Classification
    • Input data
    • Preprocessing
    • Extract feature
    • Classification model
  • ML/DL training steps:
    • Feed forward
    • Error measurement
    • Back propagation
  • RADAR
    • RAdio Detection And Ranging
    • Uses Doppler effect to measure speed directly while other sensors calculate speed based on two readings.
    • Because Radar waves bounces off hard surfaces, they can provide measurements to objects without direct line of flight.
    • Low resolution.
    • Not sensitive to weather.
  • LiDAR
    • Light Detection And Ranging
    • Uses an infrared laser beam (wave length about 900 nm).
    • Has a much higher spatial resolution than Radar.

Prediction

  • Model-based Prediction
    • Analytical based on equations.
  • Data-driven Prediction
    • More data, more accurate.
  • Apollo: Lane-sequence-based prediction
    • Predict the probability of the obstacle to follow a lane sequence based on RNNs trained from lane sequences and obstacle status.

Planning

  • Overall planning structure

    (map, start, goal)    -->   Route Navigation    -->   High level route planning (A*)
                                  |
                                  V
                            Trajectory planning -->   A sequence of collision free points

  • Trajectory generation needs to consider:

    • Collision free
    • Passenger comfort
    • Trajectory physically viable for the car
    • Obey laws

  • Use cost function to choose the best trajectory

    • Deviation from the center of lane
    • Collisions
    • Speed limit
    • Passenger comfort

  • Path-Velocity decoupled planning

    • Path planning + speed planning
      • Path planning: segment the road into cells, randomly sample points from each of these cells. Connect these points to create candidate paths, evaluate paths using cost functions.
      • speed planning: ST-graph, S: longitudinal distance. Use quadratic programming to smooth the path and speed profile.

  • Lattice Planning

    • Lateral offset relative to the longitudinal trajectory: S-L graph
    • S-T graph
    • Generate ST and SL trajectories independently and then combining them
    • ending states:
      • cruising
      • following
      • stopping
    • Once we have ST and SL trajectories, we can transform them back to the Cartesian coordinate frame. Then we can combine them to construct a 3D trajectory composed of 2D way points and 1D time stamps.

Control

Use variable control inputs to minimize deviation from target trajectory and maximize passenger comfort.

  • PID

    • equation:
      • img
    • Disadvantages
      • Just a linear algorithm, insufficient for complex problems.
      • For self driving car, we need different PID controllers for steering and acceleration, it is hard to combine a latitudinal and a longitudinal control.
      • PID depends on real-time error measurement, it may fail when it subjects to measurement delays.

  • LQR

    • Linear quadratic regulator
    • Model-based controller uses the state of the vehicle to minimize error.
    • Used for lateral control which has four inputs:
      • lateral error
      • rate of change of lateral error
      • heading error
      • rate of change of heading error
    • state - img
    • control input
      • img
    • Model
      • img
    • Cost
      • img
      • will be something like:
        • img
    • Objective - img - find optimal K

  • MPC

    • Build a model of the vehicle.
    • Use an optimization engine to calculate control inputs over a finite time horizon.
    • Implement the first set of control inputs.
    • Need to decide on how far into the future we want MPC to look.
      • Need to tradeoff accuracy with how quickly we need to get a result.
    • (Vehicle state + Control inputs) --> Vehicle model --> Vehicle Trajectory
    • Advantage:
      • Takes vehicle model, more accurate than PID
      • Work with different cost functions
    • Disadvantage:
      • Complex, slower, harder to implement

My notes when reading Apollo code

Localization

  • orientation is a quaternion that represents the rotation from the IMU coordinate (Right/Forward/Up) to the world coordinate (East/North/Up).
  • heading is zero when the car is facing East, and positive when facing North.

Control

  • What is steer_ratio?

    • The steering ratio is the ratio of the number of degrees of turn of the steering wheel to the number of degrees the wheel(s) turn as a result. In motorcycles, delta tricycles and bicycles, the steering ratio is always 1:1, because the steering wheel is fixed to the front wheel. A steering ratio of x:y means that a turn of the steering wheel x degree(s) causes the wheel(s) to turn y degree(s). In most passenger cars, the ratio is between 12:1 and 20:1.

  • steer_single_direction_max_degree
    • It is the maximum turn of steering wheel, not the maximum turn degree.
    • about 470 in apollo configuration file

Currently, two strategies for controllers in Apollo:

  • Longitudinal controller (PID) + lateral controller (LQR)
    • PID outputs throttle and brake
    • LQR output steering
  • MPC controller
    • Build a model of the vehicle
    • Use an optimization engine to calculate control inputs over a finite time horizon
    • Implement the first set of control inputs