Calibration of unmanned autonomous logistics vehicle

Calibration of autonomous vehicles is a crucial step in autonomous driving and robotics technology. Simply put, calibration involves establishing geometric and mathematical relationships between various sensors of the autonomous vehicle, as well as between the sensors and the vehicle's body coordinate system. If the calibration is inaccurate, the vehicle may "see things askew" or "deviate from the path", leading to navigation failures or even collisions.

The calibration of autonomous vehicles primarily falls into three categories: sensor intrinsic calibration, sensor extrinsic calibration (joint calibration), and chassis/kinematics calibration.
1. Sensor intrinsic parameter calibration
Internal reference calibration addresses the issue of sensor accuracy, and it is typically completed independently before shipment or installation.
①The purpose of camera intrinsic parameter calibration is to obtain the focal length, principal point, and distortion parameters (to eliminate barrel and pincushion distortions of fisheye or regular lenses) of the camera.

Common method: Zhang Zhengyou calibration method (using black and white checkerboard or dot calibration board, taking multiple photos from different angles).

②The purpose of IMU (Inertial Measurement Unit) internal parameter calibration is to determine the zero bias, scale factor, and noise of the accelerometer and gyroscope.

Common methods: static Allen variance analysis, multi-position rotation test.

2. Sensor external parameter calibration

Extrinsic calibration addresses the positional relationships between sensors and between sensors and the vehicle body (spatial relative rotation matrix R and translation vector T), which forms the foundation of multi-sensor fusion.

①Camera and LiDAR

 Objective: To align the 3D point cloud from the LiDAR with the 2D image pixels from the camera, enabling "coloring the point cloud" or "adding depth to the image".

 Method: Target calibration: Place a calibration board in the scene, simultaneously extract the checkerboard corners from the image and the edge planes from the point cloud, and solve for the extrinsic parameters using an optimization algorithm.

Target-free (online) calibration: Utilizing edges in the natural environment, such as building lines, to match the edges of image gradients and point cloud intensities.

② LiDAR and inertial navigation system/vehicle body

 Purpose: Convert the data measured by the lidar to the center of the car (usually the center of the rear axle or the IMU center) to ensure that the map building and positioning do not drift.

 Method: The car performs an "8"-shaped circular motion, straight-line driving, or in-place rotation on the horizon or within a specific field. By matching point cloud features with the integrated trajectory of the IMU, the relative pose is calculated.

③ Multi-LIDAR calibration

 Objective: When a vehicle is equipped with multiple lidars (such as one primary lidar and two blind-spot compensation lidars), the point clouds from these lidars should be stitched together to form a complete 360° view.

 Method: Extract common features (such as ground and wall surfaces) in the overlapping area and use the ICP (Iterative Closest Points) algorithm for registration.

3. Chassis and kinematics calibration

Chassis calibration addresses the accuracy issues of the actuators and odometers, preventing the robot from "thinking it has traveled one meter when it has actually only moved half a meter.".

 Calibration of wheel diameter and track: Calibrate the actual effective radius r of the two driving wheels and the actual spacing W between the wheels on both sides.

 Odometer and IMU/Laser External Reference

Method: Let the car travel straight for a certain distance or rotate 360° in place, compare the data recorded by the odometer with the absolute displacement and angle measured by the LiDAR/inertial navigation system, and correct the scale factor of the odometer.

4. Common open-source calibration toolchains

In practical development, it is generally not necessary to write algorithms from scratch, as there are a large number of mature open-source tools available in the industry:

5. Tips for implementing calibration

Rigid fixation: The sensor installation must be absolutely rigid. If minor vibrations during the trolley's travel cause the sensor to shift, the calibration will become invalid.

Inside first, then outside: It is necessary to ensure the accuracy of the intrinsic parameters of individual sensors before proceeding with the joint calibration of the extrinsic parameters between sensors.

Environment selection: For static calibration, it is recommended to choose indoor or enclosed environments with uniform lighting and rich geometric features (such as walls and pillars). For dynamic calibration, an environment with good GNSS (GPS) signals and minimal dynamic interference (such as pedestrians and passing vehicles) is required.