Multiple sensor calibration in autonomous vehicles performed in an undefined environment
The subject technology is related to autonomous vehicles (AV) and, in particular, to calibrating multiple sensors of an AV in an undefined training area. An example method includes instructing the AV to pilot itself in an undefined training area subject to at least one constraint, wherein the AV is instructed to pilot itself along a path until the AV has overlapped at least a portion of the path, and at a location at which the AV has overlapped at least the portion of the path, determining that first returns from a previous LIDAR scan overlaps with second returns from a subsequent LIDAR scan taken when the AV has overlapped at least the portion of the path. Initially, the AV does not include a location reference to identify locations of objects in the undefined training area and a plurality of sensors of the AV are uncalibrated with respect to each other.
1. A method of calibrating an autonomous vehicle (AV) in an undefined environment, the method comprising:
controlling the AV to pilot itself in an undefined training area subject to at least one constraint, wherein the undefined training area includes at least one undefined feature and wherein the AV is instructed to pilot itself along a path until the AV has overlapped at least a portion of the path, wherein the AV pilots itself through the undefined training area without reference to any previously stored map data, GNSS landmark database, or external calibration target and the at the path is autonomously generated by the AV;
at a location at which the AV has overlapped at least the portion of the path, determining that first returns from a previous LIDAR scan overlaps with second returns from a subsequent LIDAR scan taken when the AV has overlapped at least the portion of the path, wherein determining the overlap comprises:
identifying objects in the environment that were identified before performing a loop movement and identifying those objects after performing the loop movement to determine whether any of the objects are in different locations;
using multiple objects to triangulate a position of the AV and determine distances to the objects within the environment; and
validating calibration of various sensors based on comparing initial sensor measurements at the location before the loop movement with subsequent sensor measurements after returning to the location; and
executing, a three-dimensional point-cloud registration algorithm that aligns the first returns and the second returns using an iterative closest point (ICP) method to generate a six-degree-of-freedom (6-DOF) transform between the scans, and computing a root-mean-square (RMS) error between the aligned point clouds, wherein calibration is validated only when the RMS error is below a predetermined threshold of less than 5 centimeters; and
based on a determination that the first returns from the previous LIDAR scan overlaps with second returns from the subsequent LIDAR scan, determining that the AV satisfies an open road requirement to navigate an open road.
2. The method of claim 1 , further comprising:
while the AV is stationary in the undefined training area, determining whether the AV is safe to move based on a determination that a plurality of sensors of the AV are operational and providing expected data.
3. The method of claim 2 , wherein the determining of the plurality of sensors of the AV are providing the expected data comprises:
comparing first sensor data from a first sensor with second sensor data from a second sensor;
identifying a common feature represented in the first sensor data and the second sensor data; and
determining that a position of the common feature in the first sensor data corresponds to a position of the common feature in the second sensor data.
4. The method of claim 3 , further comprising:
after determining the AV is safe to move, instructing the AV to traverse a linear path in the undefined training area; and
calibrating an initial sensor of the plurality of sensors based on sensor data measured by the initial sensor during the traversal of the linear path.
5. The method of claim 4 , wherein calibrating parameters associated with at least the initial sensor comprises:
identifying features of a static object in the undefined training area using the initial sensor of the plurality of sensors; and
measuring a first odometry relative to the static object before and after traversal of the linear path using the initial sensor.
6. The method of claim 5 , further comprising:
determining a second odometry using sensor data captured by a different sensor;
comparing the first odometry to the second odometry, and
calibrating the initial sensor and the different sensor based on the comparison of the first odometry to the second odometry.
7. The method of claim 4 , further comprising:
when the AV successfully traverses the linear path, instructing the AV to traverse a non-linear path in the undefined training area; and
calibrating the first sensor and the second sensor of the plurality of sensors based on sensor data measured by the first sensor and the second sensor during traversal of the non-linear path.
8. The method of claim 7 , wherein, after the AV successfully traverses the non-linear path, the AV is instructed to pilot itself until the AV has overlapped at least the portion of the path.
9. The method of claim 8 , wherein the first sensor comprises a first light detection and ranging (LIDAR) sensor and the second sensor comprises one of a second LIDAR sensor, a radar sensor, and a time-of-flight (ToF) sensor.
10. The method of claim 1 , wherein the AV determines the path from a current position to the location and autonomously moves along the path in the undefined training area.
11. The method of claim 1 , wherein the AV is provided waypoints that identify at least a portion of the path within the undefined training area.
12. The method of claim 1 , wherein the undefined training area is located in a parking lot.
13. An autonomous vehicle (AV) comprising:
a storage configured to store instructions;
an electric drive train configured to move the AV;
a processor configured to execute the instructions and cause the processor to:
control the AV to pilot itself in an undefined training area subject to at least one constraint, wherein the undefined training area includes at least one undefined feature and wherein the AV is instructed to pilot itself along a path until the AV has overlapped at least a portion of the path, wherein the AV pilots itself through the undefined training area without reference to any previously stored map data, GNSS landmark database, or external calibration target and the at the path is autonomously generated by the AV;
at a location at which the AV has overlapped at least the portion of the path, determine that first returns from a previous LIDAR scan overlaps with second returns from a subsequent LIDAR scan taken when the AV has overlapped at least the portion of the path, wherein determining the overlap comprises:
identifying objects in the environment that were identified before performing a loop movement and identifying those objects after performing the loop movement to determine whether any of the objects are in different locations;
using multiple objects to triangulate a position of the AV and determine distances to the objects within the environment; and
validating calibration of various sensors based on comparing initial sensor measurements at the location before the loop movement with subsequent sensor measurements after returning to the location; and
executing, a three-dimensional point-cloud registration algorithm that aligns the first returns and the second returns using an iterative closest point (ICP) method to generate a six-degree-of-freedom (6-DOF) transform between the scans, and computing a root-mean-square (RMS) error between the aligned point clouds, wherein calibration is validated only when the RMS error is below a predetermined threshold of less than 5 centimeters; and
based on a determination that the first returns from the previous LIDAR scan overlaps with second returns from the subsequent LIDAR scan, determining that the AV satisfies an open road requirement to navigate an open road.
14. The AV of claim 13 , wherein the AV does not include a location reference to identify locations of existing objects in the undefined training area and a plurality of sensors of the AV are uncalibrated with respect to each other.
15. The AV of claim 14 , wherein while the AV is stationary in the undefined training area, determine whether the AV is safe to move based on the plurality of sensors of the AV are operational and providing expected data.
16. The AV of claim 15 , wherein the processor is configured to execute the instructions and cause the processor to:
compare first sensor data from a first sensor with second sensor data from a second sensor;
identify a common feature represented in the first sensor data and the second sensor data; and
determining that a position of the common feature in the first sensor data corresponds to a position of the common feature in the second sensor data.
17. The AV of claim 16 , wherein the processor is configured to execute the instructions and cause the processor to:
when the AV successfully traverses the path, instruct the AV to traverse a linear path in the undefined training area; and
calibrate an initial sensor of the plurality of sensors based on sensor data measured by the initial sensor during the traversal of the linear path.
18. The AV of claim 17 , wherein the processor is configured to execute the instructions and cause the processor to:
identify features of a static object in the undefined training area using the initial sensor of the plurality of sensors; and
measure a first odometry relative to the static object before and after traversal of the linear path using the initial sensor.
19. The autonomous vehicle of claim 18 , wherein the processor is configured to execute the instructions and cause the processor to:
determine a second odometry using sensor data captured by a different sensor;
compare the first odometry to the second odometry, and
calibrate the initial sensor and the different sensor based on the comparison of the first odometry to the second odometry.