The third lesson from the start, and finally to the very core of courses of it!
1. Positioning Introduction
Positioning is to let no one know the exact location of the car's approach. I am sure you would not want to sit on a do not know where the car now 23333!
GPS positioning accuracy is usually 1 ~ 3m, which for unmanned vehicles is not enough.
And if surrounded by tall buildings, canyons, mountains, the GPS accuracy will be even worse: it is usually in the 10 ~ 50m.
So we can not rely on GPS to determine location information, we need a more precise method. The usual practice is to see the content of automotive sensors are compared (high-precision map of lesson content) and content on the map. A distance sensor of the vehicle between the vehicle and the static obstacle (such as trees, road signs, utility poles, walls) can be measured.
We measured the distance and direction of these static obstacles in the vehicle's own coordinate system in the vehicle coordinate system, the forward direction of the car is always forward. When the car turn left or right, the coordinate system also will be rotated so that the car can always forward direction.
However, with the rotation of the car, the vehicle coordinate system is bound to be inconsistent with the map coordinates. And the vehicle coordinate system may depend on the map coordinate system set in the navigation system, it is to be understood that the switching of the coordinate system (ROS The typical conversion tf).
When a sensor to find the signs on the map, we will match the coordinates of the sensor needs to convert the measured value to the value in the map coordinates, and vice versa, which is positioned to solve the problem in a very important step.
Finally, the entire system must be under precise positioning accuracy of 10cm. There are many ways to locate the following, we will together explore the pros and cons of these methods.
2. Sebastian introduce positioning
The main areas of research is the positioning of Sebastian, has made during his tenure professor at Stanford over 100 papers on the positioning of worship big brother orz.
3. GNSS RTK_a
If you get lost in the wild, seeing a tree, far away from their own 75m, it can be determined that the tree himself on a circle of radius 75m, but can not determine the specific location:
Now, we know a house, away from their own 64m away, then we can determine their own at one of two intersections of two circles, but still do not know the specific location:
but if they know their own street a distance of 55m, now combined with maps, we can exact They know their own position:
more than a process called " triangulation " is the working principle of GPS.
4. GNSS RTK_b
GNSS: Global Navigation Satellite System (GNSS)
the GPS:, Ltd. Free Join Positioning System (GPS)
the GPS satellite navigation system is the United States. Also, Russia's GLONASS; the EU's Galileo; China's Beidou. This major navigation systems collectively referred to as GNSS
the GPS is the most widely used GNSS system. GPS is divided into three parts:
- The first part is a satellite
- The second part of the control stations located around the world
- The third part is a receiving system (e.g. mobile phones, computers and other devices)
Since the distance measured is the GPS time * speed of light, it is equipped with atomic clocks in time of the satellite, there is still an error in the distance measurement.
To be more precise, when used in conjunction with GPS usually RTK (Real-time kinematic) used, the RTK by the phase difference, the GPS positioning error can be eliminated to a large extent.
The combination of RTK GPS accuracy can be controlled within 10cm, but in the face of high-rise buildings, etc., GPS is still stretched, another disadvantage of GPS is that the update frequency is too low, can not meet the real-time requirements.
The inertial navigation _a
An inertial navigation system (the INS, INS for short) is a use of accelerometers and gyroscopes to measure the object acceleration and angular velocity , and the computer assisted navigation system to estimate the continuous moving object position, velocity and attitude. It does not require an external reference frame.
6. inertial navigation _b
INS data update rate, good short-term accuracy and stability; however, the navigation information is generated through integration because positioning error increases with time, the long-term accuracy, there will be cumulative error;
Thus, the inertial measurement unit (English: Inertial measurement unit, referred IMU) and GPS combined, make up their own disadvantages: on the one hand, the high update frequency imu make up a low update frequency gps disadvantage, on the other hand, gps corrects the motion error imu.
But in the face of such a situation, such as tunnels, underground garage, even gps + imu model, due to the long time of update gps, unmanned vehicles can not obtain accurate positioning.
7. The laser radar positioning
Lidar, may be positioned to match the point cloud unmanned vehicle. The method utilizes laser radar detection data continuously matches the pre-existing high accuracy maps, there are many algorithms for matching.
ICP (Iterative Closest Point iterative closest point algorithm), the essence of ICP algorithm is based on the optimal matching of the least squares method, it is repeated, "to determine the correspondence between the set point → calculate the optimal rigid transformation" process until a proper representation convergence criteria are met match. About ICP algorithm can refer to this blog.
Another matched filter algorithm is a point cloud algorithm. Apollo uses a histogram filtering algorithm.
Kalman filtering is another LiDAR localization algorithm, on the Kalman filter, can refer to previous blog, I have mentioned.
LiDAR location advantage is its robustness (robustness) good sex. Drawback is that there needs to advance high-precision maps and expensive laser radar, and high-precision maps require frequent updating, but this is almost impossible, because the surrounding environment is constantly changing forever. In fact, it plainly is the cost.
8. visual positioning
Only difficult to locate by the image, typically it requires integration with other sensors.
Visual positioning principle is the particle filter. The so-called particle filter refers to: by looking for a set of random samples propagated in the state space to approximate representation of the probability density function, with a sample average in place of the integration operation, and further the process of obtaining minimum variance system state estimation, these samples are images of known as "particles", therefore called particle filter.
Using particle filtering principle, a lane line camera, in order to achieve targeting to match.
Advantage is that visual positioning of low cost, easily obtained image information; disadvantage is the lack of dependence on the three-dimensional information and three-dimensional map.
9. Apollo Location
Apollo positioning based on multi-sensor uses a GPS, IMU and laser radar fusion positioning system. This fusion method utilizes the complementary strengths of different sensors to improve the accuracy and stability.
Apollo positioning module depends on the IMU, GPS, laser, radar and high-precision map. These sensors GNSS fix and support LiDAR positioning. GNSS positioning position and velocity information output, the output of the LiDAR positioning position information and the traveling direction, the frame fusion to combine these outputs Kalman filter.
Kalman filter prediction based on a two-step measurement period, the Apollo, the inertial navigation solution for the Kalman filter prediction step, for positioning the GNSS measurements and LiDAR Kalman filter update step.
10. Project example: car being targeted curriculum review & 11
slightly