Lane detection and tracking techniques for imaging systems
A method for tracking a lane on a road is presented. The method comprises receiving, by one or more processors from an imaging system, a set of pixels associated with lane markings. The method further includes generating, by the one or more processors, a predicted spline comprising (i) a first spline and (ii) a predicted extension of the first spline in a direction in which the imaging system is moving. The first spline describes a boundary of a lane and is generated based on the set of pixels. The predicted extension of the first spline is generated based at least in part on a curvature of at least a portion of the first spline.
1. A method for tracking a lane on a road, the method comprising:
receiving, by one or more processors from an imaging system, a set of pixels associated with lane markings; and
generating, by the one or more processors, a predicted spline comprising (i) a first spline and (ii) a predicted extension of the first spline in a direction in which the imaging system is moving, wherein:
the first spline describes a boundary of the lane and is generated based on the set of pixels; and
the predicted extension of the first spline is generated based at least in part on a curvature of at least a portion of the first spline.
2. The method of claim 1 , wherein generating the predicted extension of the first spline is further based on map data comprising a shape, a direction, or a curvature of the road.
3. The method of claim 1 , wherein the set of pixels is a first set of pixels generated at a first time for a first portion of the road, the method further comprising:
receiving, by the one or more processors from the imaging system, a second set of pixels associated with lane markings;
generating, by the one or more processors, an updated spline describing the boundary of the lane; and
comparing the predicted spline to the second set of pixels to determine a likelihood score.
4. The method of claim 3 , wherein comparing the predicted spline to the second set of pixels to determine the likelihood score comprises:
convolving respective values of (i) a Gaussian kernel derived from the predicted spline and (ii) a histogram representing the second set of pixels, wherein the convolving generates a set of convolved values; and
summing over the set of convolved values to determine the likelihood score.
5. The method of claim 4 , wherein each Gaussian kernel is normalized to an area.
6. The method of claim 3 , wherein the likelihood score is further based on a measure of drivability associated with the predicted spline, the measure of drivability being based on a curvature of at least a portion of the predicted spline.
7. The method of claim 3 , further comprising:
training a machine learning model to generate predicted extensions of splines using (i) a prior set of pixels associated with lane markings, (ii) a prior set of first splines based on the prior set of pixels, and (iii) a prior set of updated splines; and
applying the machine learning model to the set of pixels to generate the predicted extension of the first spline.
8. The method of claim 7 , wherein applying the machine learning model includes applying, as feature labels, indications from an operator regarding pixels in an image corresponding to the lane markings.
9. The method of claim 1 , wherein generating the predicted spline includes generating a plurality of predicted splines, each predicted spline including a first spline and a predicted extension of the first spline in the direction in which the imaging system is moving, based at least in part on the curvature of at least the portion of the first spline.
10. The method of claim 9 , wherein the predicted extension of the first spline is generated using a machine learning model, comparing the predicted spline to a second set of pixels includes comparing each of the plurality of predicted splines to the second set of pixels to determine a plurality of likelihood scores, and the method further comprises:
adjusting the machine learning model based on an analysis of the plurality of likelihood scores.
11. The method of claim 1 , wherein generating the predicted spline includes applying a particle-filter technique comprising:
generating a plurality of particles, each particle comprising a predicted spline that includes (i) a first spline and (ii) a predicted extension of the first spline, wherein the predicted extension of the first spline is based at least in part on the curvature of at least the portion of the first spline; and
determining a likelihood score for each of the plurality of predicted splines.
12. The method of claim 11 , wherein the particle-filter technique further comprises:
determining a location of the boundary of the lane by combining the plurality of particles according to associated likelihood scores.
13. The method of claim 12 , wherein combining the plurality of particles comprises:
determining a weighted average of the plurality of predicted splines to determine the location of the boundary of the lane, wherein each predicted spline is weighted according to its associated likelihood score.
14. The method of claim 1 , wherein generating the predicted spline includes:
generating a plurality of predicted splines, each predicted spline including (i) a first spline and (ii) a predicted extension of the first spline based at least in part on the curvature of at least the portion of the first spline; and
applying a perturbation to each of the predicted extensions by adding a particular noise signal to each predicted extension.
15. The method of claim 14 , wherein adding the particular noise signal comprises increasing an amount of curvature of the predicted extension for a first portion of the predicted splines and decreasing the amount of curvature of the predicted extension for a second portion of the predicted splines.
16. The method of claim 14 , wherein adding the particular noise signal comprises increasing or decreasing an amount of curvature of the predicted extension for a first portion of the predicted splines based on map data comprising a shape, a direction, or a curvature of the road.
17. The method of claim 14 , wherein the plurality of predicted splines includes at least fifty predicted splines.
18. The method of claim 1 , further comprising:
partitioning, by the one or more processors, the set of pixels into a plurality of groups, each of the plurality of groups associated with one or more control points; and
generating, by the one or more processors, a spline that traverses the control points of the plurality of groups, the spline describing a boundary of a lane.
19. The method of claim 18 , wherein generating the spline further comprises executing, by the one or more processors, a spline technique, and wherein the spline technique utilizes only the one or more control points.
20. The method of claim 19 , wherein the spline technique is a Catmull-Rom spline technique.