Generating scan patterns using cognitive lidar
A method for determining a scan pattern according to which a sensor equipped with a scanner scans a field of regard (FOR) is presented. The method comprises obtaining, by processing hardware, a plurality of objective functions, each of the objective functions specifying a cost for a respective property of the scan pattern, expressed in terms of one or more operational parameters of the scanner. The method further includes applying, by the processing hardware, an optimization scheme to the plurality of objective functions to generate the scan pattern. The method further includes scanning the FOR according to the generated scan pattern.
1. A method of determining a scan pattern according to which a sensor equipped with a scanner scans a field of regard (FOR), the method comprising:
obtaining, by processing hardware, a plurality of objective functions, each of the objective functions specifying a cost for a respective property of the scan pattern, expressed in terms of one or more operational parameters of the scanner, wherein the plurality of objective functions includes a velocity objective function specifying the cost for velocity at which the scanner scans a scan dimension of the FOR, and the velocity objective function (i) generates higher cost for higher velocity and (ii) calculates the cost of velocity based on a slope of a scanning segment and a length of the scanning segment, the slope corresponding to a displacement along the scan dimension over time;
applying, by the processing hardware, an optimization scheme to the plurality of objective functions to generate the scan pattern; and
scanning the FOR according to the generated scan pattern.
2. The method of claim 1 , wherein the velocity objective function calculates the cost of velocity (Cost v ) according to:
Cost v =(the slope of the scanning segment) 2 *the length of the scanning segment.
3. The method of claim 1 , wherein:
the plurality of objective functions includes an acceleration objective function specifying the cost for acceleration at which the scanner changes velocity of scanning a scan dimension of the FOR, and
the acceleration objective function generates higher cost for higher acceleration.
4. The method of claim 3 , wherein the acceleration objective function calculates the cost of acceleration based on a difference between slopes of two scanning segments with a shared endpoint, the slope of each scanning segment corresponding to a respective displacement along the scan dimension divided by time.
5. The method of claim 1 , wherein:
the plurality of objective functions includes a coverage objective function specifying the cost for coverage of an area covered by the scan pattern, and
the coverage objective function generates lower cost for higher coverage.
6. The method of claim 5 , wherein the coverage objective function calculates the cost of coverage based on a difference between a given coverage and a maximum coverage available to the scanner.
7. The method of claim 1 , wherein:
the plurality of objective functions includes a rescan objective function specifying the cost for rescanning a region covered by the scan pattern, and
the rescan objective function generates lower cost for a larger rescan region.
8. The method of claim 7 , wherein the rescan objective function calculates the cost of rescanning the region based on a size of the region and a time interval between re-scans of the region.
9. The method of claim 1 , wherein:
the plurality of objective functions includes a region-of-interest objective function specifying the cost for rescanning a region of interest covered by the scan pattern, and
the region-of-interest objective function generates lower cost for a larger region-of-interest region.
10. The method of claim 9 , wherein the region-of-interest objective function calculates the cost of scanning the region of interest based on a size of the region of interest relative to an overall size of the FOR.
11. The method of claim 1 , wherein:
the plurality of objective functions includes a drop-line objective function specifying the cost for dropping a scan line during a scan, and
the drop-line objective function generates higher cost for a larger number of dropped scan lines.
12. The method of claim 11 , wherein the drop-line objective function calculates the cost of dropping the scan line based on an empirically determined relationship between acceleration of the scanner and a number of dropped scan lines.
13. The method of claim 1 , wherein applying the optimization scheme includes applying a velocity constraint that specifies a maximum velocity at which the scanner is capable of scanning the FOR.
14. The method of claim 1 , wherein applying the optimization scheme includes applying an acceleration constraint that specifies a maximum acceleration with which the scanner is capable of scanning the FOR.
15. The method of claim 1 , further comprising:
applying, by the processing hardware, a reinforcement learning model to determine a lower boundary and an upper boundary for the one or more operational parameters of the scanner;
wherein applying the optimization scheme includes using the determined lower boundary and the upper boundary as optimization constraints.
16. The method of claim 15 , wherein applying the reinforcement learning model includes iteratively performing:
determining, at a first time, a first world state by scanning the FOR in accordance with a first iteration of the scan pattern;
applying a policy of the reinforcement learning model based on the first world state to generate the lower boundary and the upper boundary; and
determining, at a second time, a second world state by scanning the FOR in accordance with a second iteration of the scan pattern generated by applying the optimization scheme and using the lower boundary and the upper boundary as optimization constraints.
17. The method of claim 15 , further comprising training the reinforcement learning model by:
determining, at a first time, a first world state by scanning the FOR in accordance with a first iteration of the scan pattern;
applying a reinforcement learning policy based on the first world state to generate the lower boundary and the upper boundary;
determining, at a second time, a second world state by scanning the FOR in accordance with a second iteration of the scan pattern generated by applying the optimization scheme and using the lower boundary and the upper boundary as optimization constraints; and
determining a reward for the reinforcement learning model based on an evolution of the first world state to the second world state.
18. The method of claim 15 , wherein applying the reinforcement learning model includes determining world states at respective different times using a lidar sensor that generates point clouds at the respective different times.
19. A lidar system comprising:
a light source configured to emit light pulses;
a scanner configured to scan a field of regard (FOR) of the lidar system including direct the light pulses at different angles toward different points within the FOR; and
a controller configured to drive the scanner to direct the light pulses across the field of regard, including:
obtain a plurality of objective functions, each of the objective functions specifying a cost for a respective property of a scan pattern, expressed in terms of one or more operational parameters of the scanner, wherein the plurality of objective functions includes a velocity objective function specifying the cost for velocity at which the scanner scans a scan dimension of the FOR, and the velocity objective function (i) generates higher cost for higher velocity and (ii) calculates the cost of velocity based on a slope of a scanning segment and a length of the scanning segment, the slope corresponding to a displacement along the scan dimension over time
apply an optimization scheme to the plurality of objective functions to generate the scan pattern, and
scan the FOR according to the generated scan pattern.