Method for Managing Tracklets in a Particle Filter Estimation Framework

Abstract

A method for managing tracklets in a particle filter estimation framework includes executing a tracklet prediction dependent on a list of previous tracklets, thereby determining persistent tracklets and new tracklets; sampling new measurements for initializing the new tracklets, thereby determining an amount of estimated new tracklets; and determining an amount of the persistent tracklets dependent on the list of previous tracklets. The method further includes determining an amount of the new tracklets and an amount of updated persistent tracklets to be sampled dependent on the amount of estimated new tracklets, the amount of the persistent tracklets, and a memory bound; sampling the updated persistent tracklets from a list of the persistent tracklets dependent on the determined amount of the updated persistent tracklets; and sampling the new tracklets from unassociated measurements dependent on the determined amount of the new tracklets.

Claims

1. A method for managing tracklets in a particle filter estimation framework, comprising: executing a tracklet prediction dependent on a list of previous tracklets, thereby determining persistent tracklets and new tracklets; sampling new measurements for initializing the new tracklets, thereby determining an amount of estimated new tracklets; determining an amount of the persistent tracklets dependent on the list of previous tracklets; determining an amount of the new tracklets and an amount of updated persistent tracklets to be sampled dependent on the amount of estimated new tracklets, the amount of the persistent tracklets, and a memory bound; sampling the updated persistent tracklets from a list of the persistent tracklets dependent on the determined amount of the updated persistent tracklets; and sampling the new tracklets from unassociated measurements dependent on the determined amount of the new tracklets.

2. The method of claim 1, further comprising: validating the persistent tracklets dependent on a validation metric, thereby determining invalid tracklets; and removing the invalid tracklets from the list of previous tracklets.

3. The method of claim 2, wherein the validation metric comprises an age of a tracklet and/or an existence probability of a tracklet.

4. The method of claim 1, wherein: sampling the updated persistent tracklets comprises sampling new tracklet indices from the list of previous tracklets; and sampling the new tracklets comprises sampling new measurement indices for initializing the new tracklets.

5. The method of claim 1, wherein executing a tracklet prediction comprises determining tracklet weights of the persistent tracklets and the new tracklets.

6. The method of claim 1, wherein each tracklet is an individual particle filter-based estimator indicating a group of particles.

7. The method of claim 1, wherein determining the amount of the new tracklets comprises: determining an amount of empty slots; subtracting the amount of the persistent tracklets from the memory bound; and choosing a lower amount of the amount of estimated new tracklets and a higher amount of a maximum amount of the new tracklets and the amount of empty slots.

8. The method of claim 1, wherein determining the amount of updated persistent tracklets comprises: choosing a lower amount of the amount of the persistent tracklets and a difference between the memory bound and the amount of the new tracklets.

9. The method of claim 1, wherein a control unit is configured to execute the method.

10. A method for classifying objects in image data, comprising: receiving the image data; and classifying the objects in the image data using a particle filter estimation framework for managing tracklets, the particle filter estimation framework configured to: execute a tracklet prediction dependent on a list of previous tracklets, thereby determining persistent tracklets and new tracklets; sample new measurements for initializing the new tracklets, thereby determining an amount of estimated new tracklets; determine an amount of the persistent tracklets dependent on the list of previous tracklets; determine an amount of the new tracklets and an amount of updated persistent tracklets to be sampled dependent on the amount of estimated new tracklets, the amount of the persistent tracklets, and a memory bound; sample the updated persistent tracklets from a list of the persistent tracklets dependent on the determined amount of the updated persistent tracklets; and sample the new tracklets from unassociated measurements dependent on the determined amount of the new tracklets.

11. A control method of an at least partly autonomous robot, comprising: receiving sensor data of the at least partly autonomous robot; executing a sensor recognition on the received sensor data according to a method of managing tracklets in a particle filter estimation framework comprising: executing a tracklet prediction dependent on a list of previous tracklets, thereby determining persistent tracklets and new tracklets; sampling new measurements for initializing the new tracklets, thereby determining an amount of estimated new tracklets; determining an amount of the persistent tracklets dependent on the list of previous tracklets; determining an amount of the new tracklets and an amount of updated persistent tracklets to be sampled dependent on the amount of estimated new tracklets, the amount of the persistent tracklets, and a memory bound; sampling the updated persistent tracklets from a list of the persistent tracklets dependent on the determined amount of the updated persistent tracklets; and sampling the new tracklets from unassociated measurements dependent on the determined amount of the new tracklets; and controlling the at least partly autonomous robot dependent on the executed sensor recognition.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The subject matter of the disclosure will be explained in more detail in the following text with reference to preferred exemplary embodiments, which are illustrated in the attached drawings, in which:

[0066] FIG. 1 shows a schematic view of a relation between objects, tracklets and particles;

[0067] FIG. 2 shows a schematic view of a method for managing tracklets in a particle filter estimation framework.

[0068] FIG. 3 shows a schematic view of the particle filter estimation framework; and

[0069] FIG. 4 shows another schematic view of a method for managing tracklets in a particle filter estimation framework.

DETAILED DESCRIPTION

[0070] The reference symbols used in the drawings, and their meanings, are listed in summary form in the list of reference symbols. In principle, identical parts are provided with the same reference symbols in the figures.

[0071] FIG. 1 shows a schematic view of a relation between objects, tracklets and particles. For example a traffic scenario is observed by different sensors resulting in a grid map M of a plurality of grid cells C. In the grid map M at least one object O is identified. Particles P are grouped into tracklets T. Every tracklet T comprises a tracklet state, which is updated with an individual particle filter by employing its own particles P. Multiple particles P are used to determine objects O by clustering the particles P.

[0072] FIG. 2 shows a schematic view of a method for managing tracklets in a particle filter estimation framework.

[0073] A plurality of persistent tracklets T.sub.p,t are available in a pluarality of lists of previous tracklets h.sub.t,p at a certain timepoint t. In a next time step t+1, updated a plurality of persistent tracklets T.sub.p,t+1 are determined in a plurality of lists of updated persistent tracklets h.sub.t+1,p. The persistent tracklets T.sub.p,t+1 are either just resampled from a list of persistent tracklets h.sub.t,p or selectively resampled from particles P of persistent tracklets from a list of previous tracklets h.sub.t,p.

[0074] Additionally, lists of new tracklets h.sub.t+1,b comprising a plurality of new tracklet T.sub.new, are determined based on measurement occupancy masses m.sub.c(Occ), grid cell mass-based intensities D.sub.C and initialization weights w.sub.init.

[0075] FIG. 3 shows a schematic view of a particle filter estimation framework 10. The particle filter estimation framework 10 comprises a plurality of sensors 20, providing sensor data of for example a traffic scenario. In this case, a set of radar sensors 21, a stereo vision sensor 22 and a laser sensor 23 provide sensor data from an environment of a vehicle in the traffic scenario. For example, the sensor data cover a front view, a rear view, a left view and a right view of the vehicle. The sensor data from the stereo vision sensor 22 and the laser sensor 23 are provided to a semantic stixel unit 30, which provides a more compact Stixel-based representation of the sensor data.

[0076] The preprocessed sensor data is provided to grid channel 40, which comprise separate evidence grid channel, in this case an occupancy grid channel 41 and a semantic grid channel 42. The grid channels 40 fed with the sensor data is the basis for a multi-layer particle filter-based tracking, executed by an estimation unit 50. The estimation unit 50 provides estimated dynamic tracklets 60, which indicate a dynamic location of objects around the vehicle.

[0077] FIG. 4 shows another schematic view of a method for managing tracklets in a particle filter estimation framework. In a first step S10 a tracklet prediction is executed dependent on a list of previous tracklets T.sub.p,t, thereby determining persistent tracklets and new tracklets T.sub.new. In a second step S20, new measurements for initializing the new tracklets T.sub.new are sampled, thereby determining an amount of estimated new tracklets N.sub.eti. In a third step S30, an amount of persistent tracklets N.sub.p,t is determined dependent on the list of previous tracklets T.sub.p,t. In a fourth step S40, an amount of new tracklets N.sub.new and an amount of updated persistent tracklets N.sub.p,t+1 is determined to be sampled dependent on the amount of estimated new tracklets N.sub.eti, the amount of persistent tracklets N.sub.p,t and a memory bound N.sub.max. In a fifth step S50, updated persistent tracklets T.sub.p,t+1 are sampled from the list of persistent tracklets T.sub.p,t dependent on the determined amount of updated persistent tracklets N.sub.p,t+1. In a sixth step S60, the new tracklets T.sub.new are sampled from unassociated measurements dependent on the determined amount of new tracklets N.sub.new.