multi-object tracking with radar

23
Multi-Object Tracking with Radar Karthik Ravindran Nigam Katta

Upload: others

Post on 12-Jun-2022

12 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Multi-Object Tracking with Radar

Multi-Object Tracking with Radar

Karthik Ravindran

Nigam Katta

Page 2: Multi-Object Tracking with Radar

Agenda

2

1. Introduction to Target tracking

2. Radar sensor and what it measures

3. Kalman filter for single target tracking

4. Generalization to multiple targets

5. Addressing the Data association problem

6. Summary

Page 3: Multi-Object Tracking with Radar

Target tracking

3

Target tracking is the problem of estimating the kinematic parameters

(position, velocity etc.) of moving targets using sensor measurements

The number of targets can vary from one or to many

The sensor can itself be static or moving

Tracking is essential for environment perception in the context of autonomous

navigation

Page 4: Multi-Object Tracking with Radar

Tracking Illustration

4

Ego vehicle

sensor

Target 1

Target 2

Target 3

trajectories

Page 5: Multi-Object Tracking with Radar

Range-Bearing sensor

5

• Measures the radial distance and orientation (azimuth angle, elevation angle) of the target from the

sensor

Examples of Range-bearing sensor

(a) Radar

(b) Lidar

(c) Stereo camera

Page 6: Multi-Object Tracking with Radar

What the Radar measures

6

• In three dimensions, the sensor measures (a) range (b) azimuth (c) elevation (b) doppler

Radial velocity = doppler

Page 7: Multi-Object Tracking with Radar

What the Radar measures

7

• Multiple measurements (detections) for each target

Page 8: Multi-Object Tracking with Radar

Main components of a Tracker

8

Kalman filter

Data association

Page 9: Multi-Object Tracking with Radar

Kalman Filter

9

Uses the Range-bearing sensor measurements to estimate the positions and velocities of different targets

observed in the field of view of the sensor

Recursively estimates the kinematic parameters (position, velocity) at each time-step based on the sensor

measurements at each time-step and the previous estimates

Page 10: Multi-Object Tracking with Radar

Kalman Filter

10

Define State vector :

s = ( position, velocity, position, velocity, position, velocity … )

Define Measurement vector :

z = (range,azimuth,elevation,doppler, range,azimuth,elevation,doppler, range,azimuth,elevation,doppler … )

target 1 target 2

target 1

target 3

target 2 target 3

(position X, position Y, position Z)

(velocity X, velocity Y, velocity Z)

Page 11: Multi-Object Tracking with Radar

Measurement model

11

what we are interested in

what the sensor measures

( Measurement noise covariance matrix )

Page 12: Multi-Object Tracking with Radar

State transition model

12

( Process noise covariance matrix )

vehicle vehicle

Position 1 Position 2

(predicted)

Velocity V

Time T

Page 13: Multi-Object Tracking with Radar

Prediction-Correction steps

13

vehicle vehicle

Position 1 Position 2

(predicted)

Velocity V

Time T

vehicle

Position 2

(estimated)

vehicle

Position 2

(Radar observation)

Page 14: Multi-Object Tracking with Radar

Multiple measurements per target

14

Apply Kalman filter for each measurement (detection) and linearly combine the

individual estimates

.

.

.

.

{Detection 1}

{Detection 2}

{Detection n}

Kalman Filter 1

Kalman Filter2

Kalman Filter n

p1

p2

pn

Estimated vehicle state

Page 15: Multi-Object Tracking with Radar

Tracking multiple targets

15

• Need to identify the number of targets – Clustering

• Each cluster represents a target

• Once clustered, for every scan the detections need to be mapped

to these targets – Data association

• Every time a new target comes into the FOV of the sensor, a new

cluster is created

Page 16: Multi-Object Tracking with Radar

Data association

16

• Map the detections to different targets

• Compute an association probability each target-detection pair

target

detection

d1

d2d3

d4

d5

d6

d7

d8

Target 1

Target 2

Target 3

MulticoreWare Confidential

Page 17: Multi-Object Tracking with Radar

Clutter

17

• Spurious detections from static targets (Road, Clouds, Sea etc.)

target

detection from target

Clutter detection

MulticoreWare Confidential

Page 18: Multi-Object Tracking with Radar

Data association methods

18

• Nearest neighbourhood method

• Probabilistic data association

• Joint probabilistic data association

MulticoreWare Confidential

Page 19: Multi-Object Tracking with Radar

Nearest Neighbourhood method

19

• Detections are mapped to its nearest target

• Does not discriminate clutter

MulticoreWare Confidential

Page 20: Multi-Object Tracking with Radar

Probabilistic data association

20

• Assumes a single target in the FOV

• Assumes a probabilistic model for the spatial distribution of clutter

• Computes an association probability for each detection

• Detections probable of being a clutter will assume smaller association probabilities

MulticoreWare Confidential

Page 21: Multi-Object Tracking with Radar

Joint probabilistic data association

21

• Assumes multiple targets in the FOV

• Assumes a probabilistic model for the spatial distribution of clutter

• Computes an association probability for each detection-target pair

• Detections probable of being a clutter will assume smaller association probabilities

p1p2

p3

p4p5

p6

p7p8

p9

p1, p2, p3 ….p9are theassociationprobabilitiesconsidering theother objects inthe scenario.

MulticoreWare Confidential

Page 22: Multi-Object Tracking with Radar

Summary

22

• Defined the Target tracking problem

• Radar sensor and its measurements (detections)

• Kalman filter for state estimation

• Generalization to multiple detections, multiple targets

• Clutter detections

• Data association methods

Page 23: Multi-Object Tracking with Radar

Thank you

connect with us

www.multicorewareinc.com

www.facebook.com/multicoreware

www.twitter.com/multicoreware

www.linkedin.com/company/multicoreware-inc

www.instagram.com/multicoreware

www.youtube.com/multicoreware