Trajectories and Maneuvers of Surrounding Vehicles with Panoramic Camera Arrays

Research output: Contribution to journalJournal articleResearchpeer-review

Abstract

Vision-based research for intelligent vehicles have traditionally focused on specific regions around a vehicle, such as a front looking camera for, e.g., lane estimation. Traffic scenes are complex and vital information could be lost in unobserved regions. This paper proposes a framework that uses four visual sensors for a full surround view of a vehicle in order to achieve an understanding of surrounding vehicle behaviors. The framework will assist the analysis of naturalistic driving studies by automating the task of data reduction of the observed trajectories. To this end, trajectories are estimated using a vehicle detector together with a multiperspective optimized tracker in each view. The trajectories are transformed to a common ground plane, where they are associated between perspectives and analyzed to reveal tendencies around the ego-vehicle. The system is tested on sequences from 2.5 h of drive on US highways. The multiperspective tracker is tested in each view as well as for the ability to associate vehicles bet-ween views with a 92% recall score. A case study of vehicles approaching from the rear shows certain patterns in behavior that could potentially influence the ego-vehicle.
Original languageEnglish
JournalI E E E Intelligent Vehicles Symposium
Volume1
Issue number2
Pages (from-to)203-214
ISSN1931-0587
DOIs
Publication statusPublished - 28 Oct 2016

Fingerprint

Cameras
Trajectories
Intelligent vehicle highway systems
Data reduction
Detectors
Sensors

Cite this

@article{4c95b4e1d5b14b89823eeed452e312e0,
title = "Trajectories and Maneuvers of Surrounding Vehicles with Panoramic Camera Arrays",
abstract = "Vision-based research for intelligent vehicles have traditionally focused on specific regions around a vehicle, such as a front looking camera for, e.g., lane estimation. Traffic scenes are complex and vital information could be lost in unobserved regions. This paper proposes a framework that uses four visual sensors for a full surround view of a vehicle in order to achieve an understanding of surrounding vehicle behaviors. The framework will assist the analysis of naturalistic driving studies by automating the task of data reduction of the observed trajectories. To this end, trajectories are estimated using a vehicle detector together with a multiperspective optimized tracker in each view. The trajectories are transformed to a common ground plane, where they are associated between perspectives and analyzed to reveal tendencies around the ego-vehicle. The system is tested on sequences from 2.5 h of drive on US highways. The multiperspective tracker is tested in each view as well as for the ability to associate vehicles bet-ween views with a 92{\%} recall score. A case study of vehicles approaching from the rear shows certain patterns in behavior that could potentially influence the ego-vehicle.",
author = "Dueholm, {Jacob Velling} and Kristoffersen, {Miklas Str{\o}m} and Satzoda, {Ravi K.} and Moeslund, {Thomas B.} and Trivedi, {Mohan M.}",
year = "2016",
month = "10",
day = "28",
doi = "10.1109/TIV.2016.2622921",
language = "English",
volume = "1",
pages = "203--214",
journal = "I E E E Intelligent Vehicles Symposium",
issn = "1931-0587",
publisher = "IEEE",
number = "2",

}

Trajectories and Maneuvers of Surrounding Vehicles with Panoramic Camera Arrays. / Dueholm, Jacob Velling; Kristoffersen, Miklas Strøm; Satzoda, Ravi K.; Moeslund, Thomas B.; Trivedi, Mohan M.

In: I E E E Intelligent Vehicles Symposium, Vol. 1, No. 2, 28.10.2016, p. 203-214.

Research output: Contribution to journalJournal articleResearchpeer-review

TY - JOUR

T1 - Trajectories and Maneuvers of Surrounding Vehicles with Panoramic Camera Arrays

AU - Dueholm, Jacob Velling

AU - Kristoffersen, Miklas Strøm

AU - Satzoda, Ravi K.

AU - Moeslund, Thomas B.

AU - Trivedi, Mohan M.

PY - 2016/10/28

Y1 - 2016/10/28

N2 - Vision-based research for intelligent vehicles have traditionally focused on specific regions around a vehicle, such as a front looking camera for, e.g., lane estimation. Traffic scenes are complex and vital information could be lost in unobserved regions. This paper proposes a framework that uses four visual sensors for a full surround view of a vehicle in order to achieve an understanding of surrounding vehicle behaviors. The framework will assist the analysis of naturalistic driving studies by automating the task of data reduction of the observed trajectories. To this end, trajectories are estimated using a vehicle detector together with a multiperspective optimized tracker in each view. The trajectories are transformed to a common ground plane, where they are associated between perspectives and analyzed to reveal tendencies around the ego-vehicle. The system is tested on sequences from 2.5 h of drive on US highways. The multiperspective tracker is tested in each view as well as for the ability to associate vehicles bet-ween views with a 92% recall score. A case study of vehicles approaching from the rear shows certain patterns in behavior that could potentially influence the ego-vehicle.

AB - Vision-based research for intelligent vehicles have traditionally focused on specific regions around a vehicle, such as a front looking camera for, e.g., lane estimation. Traffic scenes are complex and vital information could be lost in unobserved regions. This paper proposes a framework that uses four visual sensors for a full surround view of a vehicle in order to achieve an understanding of surrounding vehicle behaviors. The framework will assist the analysis of naturalistic driving studies by automating the task of data reduction of the observed trajectories. To this end, trajectories are estimated using a vehicle detector together with a multiperspective optimized tracker in each view. The trajectories are transformed to a common ground plane, where they are associated between perspectives and analyzed to reveal tendencies around the ego-vehicle. The system is tested on sequences from 2.5 h of drive on US highways. The multiperspective tracker is tested in each view as well as for the ability to associate vehicles bet-ween views with a 92% recall score. A case study of vehicles approaching from the rear shows certain patterns in behavior that could potentially influence the ego-vehicle.

U2 - 10.1109/TIV.2016.2622921

DO - 10.1109/TIV.2016.2622921

M3 - Journal article

VL - 1

SP - 203

EP - 214

JO - I E E E Intelligent Vehicles Symposium

JF - I E E E Intelligent Vehicles Symposium

SN - 1931-0587

IS - 2

ER -