Stable radar track velocity initialization using multiple hypotheses
This document describes an object tracker that performs stable velocity initialization for radar tracks, using multiple hypotheses, including when only sparse radar point clouds are available. With just a single point per scan, the tracker creates multiple hypotheses for the direction and speed of an object. A least square function can be applied to each hypothesis to derive each respective initial velocity, which are tracked using a Kalman Filter during a hypotheses tracking period. When hypotheses are initialized and tracked on each hypothesis tracking period, their track error scores are computed. Based on their track error scores, the hypotheses that have low evidence are discarded during the hypotheses tracking period. When the hypotheses tracking period ends, a hypothesis with high evidence initializes the track's velocity. Parallel hypothesis evaluation enables the tracker to initialize a velocity quickly and accurately by merely selecting the best hypothesis, which may enable safer driving.
1. A device comprising a processor configured to:
obtain point cloud sensor data indicative of signal returns that reflect off of objects in an environment; and
establish, using the point cloud sensor data, a track to an object in the environment including initializing a velocity measurement for the object by:
creating multiple hypotheses for predicted movement of the object;
determining a track error score for each hypothesis of the multiple hypotheses, the track error score being computed as accumulated position and range rate errors associated with one or more points of the point cloud sensor data associated with the track;
determining, for each of the multiple hypotheses, an initial velocity and a first associated level of evidence supporting the initial velocity of that hypothesis, the first associated level of evidence corresponding to the track error score of that hypothesis;
generating a fused hypothesis by combining the multiple hypotheses, an initial velocity and a first associated level of evidence of the fused hypothesis being based on an aggregate of the initial velocities and the first associated levels of evidence of the multiple hypotheses;
including the fused hypothesis among the multiple hypotheses; and
responsive to including the fused hypothesis among the multiple hypotheses, select, based on the first associated levels of evidence for the multiple hypotheses, a first best hypothesis, the first best hypothesis having a lowest associated track error score from among the multiple hypotheses;
time update the initial velocity of each of the multiple hypotheses;
determine, for each of the multiple hypotheses, a second associated level of evidence supporting the updated initial velocity of that hypothesis, the second associated level of evidence corresponding to an updated track error score of that hypothesis;
eliminate from the multiple hypotheses any of the multiple hypotheses having values for the second associated level of evidence that do not satisfy an evidence threshold;
responsive to determining that more than two hypotheses remain in the multiple hypotheses after eliminating any of the multiple hypotheses having values for the second associated levels of evidence that do not satisfy the evidence threshold:
measurement update the predicted movement of the object for each of the remaining multiple hypotheses; and
select, based on the second associated levels of evidence for the remaining multiple hypotheses, a second best hypothesis to replace the first best hypothesis selected previously, the second best hypothesis having the lowest associated track error score of the remaining multiple hypotheses after the measurement update; and
responsive to determining that only two hypotheses remain in the multiple hypotheses after eliminating any of the multiple hypotheses having values for the second associated levels of evidence that do not satisfy the evidence threshold:
select, based on the second associated levels of evidence for the two hypotheses that remain, a third best hypothesis to replace the first or second best hypothesis selected previously, the third best hypothesis having the lowest associated track error score of the two hypotheses that remain; and
output, for a vehicle system, the track to the object including a velocity parameter initialized to the initial velocity of the third best hypothesis.
2. The device of claim 1 , wherein the processor is further configured to, until only two hypotheses remain:
time update the initial velocity of each of the multiple hypotheses;
determine, for each of the multiple hypotheses, the second associated level of evidence supporting the updated initial velocity of that hypothesis; and
eliminate from the multiple hypotheses any of the multiple hypotheses having values for the second associated level of evidence that do not satisfy the evidence threshold.
3. The device of claim 1 , wherein the second best hypothesis to replace the first best hypothesis selected previously is selected further in response to determining a hypotheses tracking period ends.
4. The device of claim 3 , wherein the hypotheses tracking period of time comprises multiple frames of a radar system from which the point cloud sensor data is obtained.
5. The device of claim 4 , wherein the hypotheses tracking period of time comprises approximately fifteen frames of the radar system.
6. The device of claim 1 , wherein the processor is configured to determine, for each of the multiple hypotheses, the first associated level of evidence supporting the initial velocity of that hypothesis by determining accumulated position and range rate errors of one or more points of the point cloud sensor data.
7. The device of claim 1 , wherein the processor is configured to use a constant motion model to time update or to measurement update the multiple hypotheses.
8. The device of claim 1 , wherein the evidence threshold comprises an evidence ratio computed for each of the multiple hypotheses.
9. The device of claim 8 , wherein the evidence ratio computed for each of the multiple hypotheses comprises a unique evidence ratio among all the multiple hypotheses.
10. The device of claim 1 , wherein the point cloud sensor data comprises point cloud radar data.
11. A method, comprising:
obtaining, by an object tracker and from a radar system, point cloud sensor data indicative of radar returns that reflect off of objects in an environment; and
establishing, using the point cloud sensor data, a track to an object in the environment including initializing a velocity measurement for the object by:
creating multiple hypotheses for predicted movement of the object;
determining a track error score for each hypothesis of the multiple hypotheses, the track error score being computed as accumulated position and range rate errors associated with one or more points of the point cloud sensor data associated with the track;
determining, for each of the multiple hypotheses, an initial velocity and a first associated level of evidence supporting the initial velocity of that hypothesis, the first associated level of evidence corresponding to the track error score of that hypothesis;
generating a fused hypothesis by combining the multiple hypotheses, an initial velocity and a first associated level of evidence of the fused hypothesis being based on an aggregate of the initial velocities and the first associated levels of evidence of the multiple hypotheses;
including the fused hypothesis among the multiple hypotheses; and
responsive to including the fused hypothesis among the multiple hypotheses, selecting, based on the first associated levels of evidence for the multiple hypotheses, a first best hypothesis for initializing the velocity measurement for the object, the first best hypothesis having a lowest associated track error score from among the multiple hypotheses.
12. The method of claim 11 , further comprising
time updating the initial velocity of each of the multiple hypotheses;
determining, for each of the multiple hypotheses, a second associated level of evidence supporting the updated initial velocity of that hypothesis, the second associated level of evidence corresponding to an updated track error score of that hypothesis;
eliminating from the multiple hypotheses any of the multiple hypotheses having values for the second associated level of evidence that do not satisfy an evidence threshold;
responsive to determining that more than two hypotheses remain in the multiple hypotheses after eliminating any of the multiple hypotheses having values for the second associated levels of evidence that do not satisfy the evidence threshold:
measurement updating the predicted movement of the object for each of the remaining multiple hypotheses; and
selecting, based on the second associated levels of evidence for the remaining multiple hypotheses, a second best hypothesis to replace the first best hypothesis selected previously, the second best hypothesis having the lowest associated track error score of the remaining multiple hypotheses after the measurement update.
13. The method of claim 11 , further comprising
time updating the initial velocity of each of the multiple hypotheses;
determining, for each of the multiple hypotheses, a second associated level of evidence supporting the updated initial velocity of that hypothesis, the second associated level of evidence corresponding to an updated track error score of that hypothesis;
eliminating from the multiple hypotheses any of the multiple hypotheses having values for the second associated level of evidence that do not satisfy an evidence threshold;
responsive to determining that only two hypotheses remain in the multiple hypotheses after eliminating any of the multiple hypotheses having values for the second associated levels of evidence that do not satisfy the evidence threshold:
selecting, based on the second associated levels of evidence for the two hypotheses that remain, a third best hypothesis to replace the first or second best hypothesis selected previously, the third best hypothesis having the lowest associated track error score of the two hypotheses that remain; and
outputting, for a vehicle system, the track to the object including a velocity parameter initialized to the initial velocity of the third best hypothesis.
14. The method of claim 11 , further comprising
time updating the initial velocity of each of the multiple hypotheses;
determining, for each of the multiple hypotheses, a second associated level of evidence supporting the updated initial velocity of that hypothesis, the second associated level of evidence corresponding to an updated track error score of that hypothesis;
eliminating from the multiple hypotheses any of the multiple hypotheses having values for the second associated level of evidence that do not satisfy an evidence threshold;
responsive to determining that more than two hypotheses remain in the multiple hypotheses after eliminating any of the multiple hypotheses having values for the second associated levels of evidence that do not satisfy the evidence threshold:
measurement updating the predicted movement of the object for each of the remaining multiple hypotheses; and
selecting, based on the second associated levels of evidence for the remaining multiple hypotheses, a second best hypothesis to replace the first best hypothesis selected previously, the second best hypothesis having the lowest associated track error score of the remaining multiple hypotheses after the measurement update; and
responsive to determining that only two hypotheses remain in the multiple hypotheses after eliminating any of the multiple hypotheses having values for the second associated levels of evidence that do not satisfy the evidence threshold:
selecting, based on the second associated levels of evidence for the two hypotheses that remain, a third best hypothesis to replace the first or second best hypothesis selected previously, the third best hypothesis having the lowest associated track error score of the two hypotheses that remain; and
outputting, for a vehicle system, the track to the object including a velocity parameter initialized to the initial velocity of the third best hypothesis.
15. The method of claim 14 , wherein the second best hypothesis to replace the first best hypothesis selected previously is selected further in response to determining a hypotheses tracking period of time is expired.
16. The method of claim 15 , wherein the hypotheses tracking period of time comprises multiple frames of the radar system.
17. The method of claim 14 , wherein the evidence threshold comprises a unique evidence ratio computed for each of the multiple hypotheses.
18. The method of claim 11 , wherein the object tracker is configured to determine, for each of the multiple hypotheses, the first associated level of evidence supporting the initial velocity of that hypothesis by determining accumulated position and range rate errors of an associated portion of the point cloud sensor data.
19. The method of claim 11 , wherein the object tracker is configured to use a constant motion model to time update or to measurement update the multiple hypotheses.
20. A non-transitory computer-readable storage medium comprising instructions that, when executed, cause a processor to execute an object tracker configured to:
obtain, from a radar system, point cloud sensor data indicative of radar returns that reflect off of objects in an environment; and
establish, using the point cloud sensor data, a track to the object in an environment including initializing a velocity measurement for the object by:
creating at least two hypotheses for predicted movement of the object;
determining a track error score for each hypothesis of the multiple hypotheses, the track error score being computed as accumulated position and range rate errors associated with one or more points of the point cloud sensor data associated with the track;
determining, for each of the at least two hypotheses, an initial velocity and a first associated level of evidence supporting the initial velocity of that hypothesis, the first associated level of evidence corresponding to the track error score of that hypothesis;
generating a fused hypothesis by combining the at least two hypotheses, an initial velocity and a first associated level of evidence of the fused hypothesis being based on an aggregate of the initial velocities and the first associated levels of evidence of the at least two hypotheses;
including the fused hypothesis among the at least two hypotheses; and
responsive to including the fused hypothesis among the at least two hypotheses, selecting, based on the first associated levels of evidence for the at least two hypotheses, a single best hypothesis for initializing the velocity measurement for the object, the first best hypothesis having a lowest associated track error score from among the multiple hypotheses.