Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Detection and tracking of Mmoving objects using stereo vision cameras
Thesis (MEng)--Stellenbosch University, 2018.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2018
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867614115808149504 |
|---|---|
| access_status_str | Open Access |
| author | Roelofse, Christiaan E. |
| author2 | Van Daalen, Corne E. |
| author_browse | Roelofse, Christiaan E. Van Daalen, Corne E. |
| author_facet | Van Daalen, Corne E. Roelofse, Christiaan E. |
| author_sort | Roelofse, Christiaan E. |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2018. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/103434 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA |
| last_indexed | 2026-06-10T12:46:55.034Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2018 |
| publishDateRange | 2018 |
| publishDateSort | 2018 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/103434 Detection and tracking of Mmoving objects using stereo vision cameras Roelofse, Christiaan E. Van Daalen, Corne E. Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering. Motion picture cameras UCTD Detection (Electronics) Automatic tracking Thesis (MEng)--Stellenbosch University, 2018. ENGLISH ABSTRACT: Detection and tracking of moving objects (DATMO) is one of the core components necessary for autonomous navigation of a robot. The robot requires measurements of its environment, and uses these measurements to construct a representation of the dynamic objects in its environment. Once dynamic objects have been detected, the robot can use their locations and movement for autonomous functionality such as navigation, or collision prediction and avoidance. Of particular interest are vision-based sensors such as cameras, due to the amount of information they give about the environment. However, the amount of information causes difficulty regarding the interpretation and workable utilisation of the information. This thesis outlines a systematic approach for robust estimation of states for DATMO using stereo vision cameras. The mathematical basis of the camera geometry is derived. These include the camera projection transform, camera parameters, and epipolar geometry. Imagefeatures are obtained from both the left and right cameras and are matched. These matched image-feature pairs are triangulated to form 3D measurements of the moving objects in the robot’s environment. These 3D measurements are then used to filter the state estimates of the moving objects. Popular feature detection algorithms are expounded and investigated. ORB, KAZE, and A-KAZE are chosen for implementation and comparison. Factors pertaining to feature detection and matching, such as subpixel accuracy and matching strength, are weighed in light of a desired robust implementation. The data association problem, that is which objects in the environment caused which measurements, is addressed. Methods used to address the problem, such as global nearest neighbour, probabilistic data association, and multiple hypothesis tracking (MHT), are examined. A multiple hypothesis tracking solution that uses Bayesian statistics is used in order to reliably associate measurements to objects in the robot’s environment. Necessary approximations to the MHT approach are made and justified. The approximations result in a first-order approximation and a Gaussian mixture density description. The issue of unbounded associations is addressed and managed with techniques that remove, approximate, or prevent unnecessary state estimates. Algorithms are tested using the KITTI dataset in Python. LiDAR is used to evaluate the results of the algorithm. The computational cost of the algorithm is the biggest issue highlighted by the results. This is due to the complexity of the multiple hypothesis tracking solution and the large number of image-features used to ensure robust and reliable functionality. The results of the thesis demonstrate that there is philosophical conflict between the requirement of robust estimation on the filtering side, and the large number of measurements required from camera images. The complexity increases with the number of measurements, but many measurements are needed in order to provide a reliable representation of the environment from camera images. AFRIKAANSE OPSOMMING: Deteksie en volging van bewegende voorwerpe (DATMO) is een van die kernkomponente wat nodig is vir outonome navigasie van ’n robot. Die robot benodig metings van die omgewing en gebruik hierdie metings om ’n voorstelling van die dinamiese voorwerpe in sy omgewing te maak. Sodra dinamiese voorwerpe gevolg is, kan die robot hul posisies en beweging gebruik vir outonome funksies soos navigasie, of botsing-voorspelling en vermyding. Van belang is visie-gebaseerde sensors soos kameras, gegee die hoeveelheid inligting wat hulle beskikbaar maak oor omgewing. Die hoeveelheid inligting veroorsaak egter probleme met betrekking tot die interpretasie en werkbare benutting van die inligting. Hierdie thesis beskryf ’n sistematiese benadering vir robuuste afskatting van toestande vir DATMO deur gebruik te maak van stereo-visie kameras. Die wiskundige basis van die kamera geometrie word afgelei. Dit sluit in die kamera projeksie transformasie, kamera parameters en epipolêre geometrie. Image-features word ontdek uit beide die linker en regter kameras en word geassosieer met mekaar. Hierdie geassosieerde image-feature-pare word gebruik om 3Dmetings van die bewegende voorwerpe in die robot se omgewing te kry. Hierdie 3D-metings word dan gebruik om die toestandafskattings van die bewegende voorwerpe te filter. Die mees populêre feature algoritmes word uiteengesit en ondersoek. ORB, KAZE en A-KAZE word gekies vir implementering en vergelyking. Faktore wat verband hou met feature-opsporing en assosiasie, soos subpixel-akkuraatheid en assosiasie betroubaarheid, word geweeg in die lig van die verlangde robuuste implementering. Die data-assosiasie probleem, dit wil sê watter voorwerpe in die omgewing veroorsaak watter metings, word aangespreek. Metodes wat gebruik word om die probleem aan te spreek word ondersoek. ’n MHT oplossing wat Bayesiese statistieke gebruik, word gebruik om metings betroubaar te assosieer met voorwerpe in die robot se omgewing. Noodsaaklike benaderings tot die MHT-benadering word gemaak en geregverdig, wat lei tot ’n eerste-orde benadering en ’n Gaussiese mengsel digtheidsbeskrywing. Die probleem van oneindige meting assosiasies word aangespreek met tegnieke wat onnodige toestandafskattings verwyder, benader of voorkom. Algoritmes word getoets met behulp van die KITTI datastel in Python. LiDAR is gebruik om die resultate van die algoritme te evalueer. Die berekeningskoste van die algoritme is die grootste probleem wat deur die resultate uitgelig word. Dit is as gevolg van die kompleksiteit van die MHT oplossing en die groot aantal image-features wat gebruik word om robuuste en betroubare funksionaliteit te verseker. Die resultate van die thesis wys dat daar filosofiese konflik is tussen die vereiste van robuuste afskatting op die filterkant en die groot aantal metings wat van kamerabeelde vereis word. Die kompleksiteit neem toe met die aantal metings, maar baie metings is benodig om ’n betroubare voorstelling van die omgewing uit kamerabeelde te verseker. 2018-02-20T09:32:32Z 2018-04-09T06:56:28Z 2018-02-20T09:32:32Z 2018-04-09T06:56:28Z 2018-03 Thesis http://hdl.handle.net/10019.1/103434 en_ZA Stellenbosch University 146 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Motion picture cameras UCTD Detection (Electronics) Automatic tracking Roelofse, Christiaan E. Detection and tracking of Mmoving objects using stereo vision cameras |
| title | Detection and tracking of Mmoving objects using stereo vision cameras |
| title_full | Detection and tracking of Mmoving objects using stereo vision cameras |
| title_fullStr | Detection and tracking of Mmoving objects using stereo vision cameras |
| title_full_unstemmed | Detection and tracking of Mmoving objects using stereo vision cameras |
| title_short | Detection and tracking of Mmoving objects using stereo vision cameras |
| title_sort | detection and tracking of mmoving objects using stereo vision cameras |
| topic | Motion picture cameras UCTD Detection (Electronics) Automatic tracking |
| url | http://hdl.handle.net/10019.1/103434 |
| work_keys_str_mv | AT roelofsechristiaane detectionandtrackingofmmovingobjectsusingstereovisioncameras |