Robotics-Robot Perception (location)

Finally finished the robotic SLAM all the content. To tell the truth, the course content is more general, but the work is quite challenging. The final chapter contains location. Location is the inverse process of Mapping. In the case of a given map, the position and posture of the robot need to be obtained.

1. The meaning of location

In the robot navigation task, location can tell the robot's current position to facilitate closed-loop control or trajectory planning. Under normal circumstances, location can be completed by Gps,wifi and so on. GPS positioning accuracy of about 3.5 meters, WiFi is more than 10 meters. For robots, unmanned vehicles, such precision is obviously unacceptable. LiDAR in the distance of 10m can achieve cm accuracy, binocular vision in 4m can achieve 10cm accuracy, compared with the GPS has a certain advantage, in addition, this non-online positioning method can be used indoors.

Location's greatest difficulty comes from measuring errors. The counting error of the odometer, the non-uniformity of the measurement and the mileage meter, etc. So we need an algorithm that can determine the position and posture of a robot with high precision.

2, location of the process

The process of location can be summarized as follows: 1, to determine the initial posture of the robot, 2, to find the sensor measurement results, 3, the use of particle filter based on the current posture and robot measurement results to estimate the next position of the robot posture.

Here the particle filter function seems similar to the Kalman filter, but the particle filter can have a non-0 mean (encoder reading), can have a non-positive distribution (how do I sample), can have a nonlinear solution (anyway, is a miracle, what solution is impossible?) ）

The initial pose of the robot is determined to be skipped. The acquisition of the sensor measurement results is also called the registration of the map. This is a word we are very familiar with--registration. The role of registration is to match the results of the measurements with the map results. The way the match is used is the factor product of the probability graph. For the positioning posture, if the measured result matches the actual map, the positive points are given, otherwise the negative points are given. As to how many positive points, how many negative points, need to carry out tuning .... This tuning is also the core of the job. The key to tuning is to visualize the measurement results, and if the first frame is not matched, then it must be run. If the first frame is matched, then there is the possibility of the adjustment behind it.

　　

3. Particle filter

The robot pose estimation based on particle filter is a large module with the following pipeline:

1. For the initial posture, given a bunch of particles, each particle is a pose, the particles can be evenly distributed, or it can be too distributed, or it can be any distribution you like. The score for each particle count registration

　　　　　　

2. Using the known transfer relation-encoder, inertial sensor, whatever, gives the next pose of the particle. Notice that the particle is transferred with its score.

3. Resampling of particles after transfer-the result of a particle transfer is resampled based on the noise of the transfer relationship. Calculates and updates the registration score.

4. Calculate the number of active particles and resample if they are less than the total number of particles. If the total number of particles is equal, return to 2.

　　

All in all, particle filters, called "multi-party politics" , are likely to be presidents. Each transfer is equivalent to an election, and the registration of the map results in the equivalent of a ballot paper. Although we will finally choose one of the most votes, but we also do not allow a party alone (n_effective). Resampling is that if the idea of a party is really good, then allow it to be split, add noise behind it to ensure the diversity of opinions.

So the particle filter always gives better results.

Robotics-Robot Perception (location)