Title Pose optimized multiple camera systems for vehicle surround-view vision Title (croatian) Sustav za panoramsku vizualizaciju iz vozila s optimalnim položajem više kamera Author Venkata Anirudh Puligandla Mentor Sven Lončarić (mentor)Committee member Marko Subašić (predsjednik povjerenstva)Committee member Vladan Papić (član povjerenstva)Committee member Tomislav Petković (član povjerenstva)Granter University of Zagreb Defense date and country 2024-04-23, Croatia Scientific / art field, TECHNICAL SCIENCES Universal decimal classificationUDC ) 004 - Computer science and technology. Computing. Data processing Abstract Advanced Driver Assistance Systems (ADAS) are ubiquitous in present day vehicles. Among various ADAS systems, parking assistance systems offering novel views of the vehicle's surroundings, such as, 360o and bird's-eye views are becoming a part of all modern vehicles. Presenting novel views involves multiple steps of image processing including, image/video capture, image registration and visualization. Capturing surrounding view requires the placement of multiple cameras on the vehicle as a first step. Precise placement and calibration of the cameras is important as minor errors may lead to significant artefacts during the subsequent steps of image registration and visualization.
Camera placement optimization (CPO) aims to optimize the poses of multiple cameras with an objective to increase the overall coverage of the target area, and/or to reduce the cost of the multiple-camera system. CPO can also eliminate the requirement for camera calibration, as the precise pose of the cameras is already estimated during the optimization step. Although CPO problems are well-studied for surveillance scenarios, there exists a dearth of literature in the context of applications to vehicle surrounding view capture. Compared to surveillance scenarios, CPO problems for vehicle surround view capture need to address additional challenges posed by the complex, non-convex structure of vehicles, and the requirement of a high degree of accuracy in the estimated camera's pose. CPO problems are simulated in discrete space by sampling the continuous space. Although modelling in discrete space is the favoured approach for their simplicity, few works use continuous space models or a mix of both for added accuracy.
The scope of this work includes CPO problem formulation for surround-view coverage for vehicles, CPO problem definitions in discrete as well as continuous space domains, and proposing a new heuristic algorithm to improve the performance of existing optimization algorithms. Firstly, new contributions are made towards formulating the CPO problem for surround-view coverage using a 3D discrete space model. A novel multi-resolution heuristic optimization algorithm is proposed to significantly improve the performance of existing discrete optimization algorithms. The CPO problem is then reformulated in the continuous space domain and compared against the discrete-space variant to highlight improved accuracy. Lastly, a super-voxel segmentation method, which was tailored for use in the \multi-resolution optimization method, is introduced and validated on well-known 3D point cloud datasets. Experiments and simulation results on high-resolution 3D models of a variety of vehicles show that the proposed methods are effective in optimizing camera poses of multiple cameras for vehicle surround-view, meeting the demands of real-world scenarios in a reasonable amount of time.
Abstract (croatian) Napredni sustavi pomoći vozaču (ADAS) su sveprisutni u današnjim vozilima. Među raznim ADAS sustavima, sustavi pomoći pri parkiranju koji nude nove poglede na okolinu vozila, kao što su 360 stupnjeva i pogled iz ptičje perspektive, postaju dio svih modernih vozila. Predstavljanje novih prikaza uključuje više koraka obrade slike uključujući snimanje slike/videozapisa, registraciju slike i vizualizaciju. Snimanje okolnog pogleda kao prvi korak zahtijeva postavljanje više kamera na vozilo. Precizno postavljanje i kalibracija kamera je važno jer manje pogreške mogu dovesti do značajnih artefakata tijekom sljedećih koraka registracije slike i vizualizacije.
Optimizacija postavljanja kamere (CPO) ima cilj optimizirati položaje više kamera s ciljem povećanja ukupne pokrivenosti ciljanog područja i/ili smanjenja troškova sustava s više kamera. CPO također može eliminirati zahtjev za kalibracijom kamere, budući da je precizna pozicija kamera već procijenjena tijekom koraka optimizacije. Iako su problemi CPO-a dobro proučeni za scenarije nadzora, postoji nedostatak literature u kontekstu primjena za snimanje pogleda okoline vozila. U usporedbi sa scenarijima nadzora, CPO problemi za snimanje vozila iz okruženja moraju se baviti dodatnim izazovima koje postavlja složena, nekonveksna struktura vozila i zahtjev visokog stupnja točnosti u procijenjenom položaju kamere. CPO problemi se simuliraju u diskretnom prostoru uzorkovanjem kontinuiranog prostora. Iako je modeliranje u diskretnom prostoru omiljeni pristup zbog svoje jednostavnosti, malo radova koristi modele kontinuiranog prostora ili njihovu kombinaciju za dodatnu točnost.
Opseg ovog rada uključuje formulaciju CPO problema za pokrivenost vozila prostornim pogledom, definicije CPO problema u diskretnim i kontinuiranim prostornim domenama i predlaganje novog heurističkog algoritma za poboljšanje performansi postojećih optimizacijskih algoritama. Prvo, dani su novi doprinosi formuliranju CPO problema za pokrivenost prostornim pogledom korištenjem 3D modela diskretnog prostora. Predložen je novi heuristički optimizacijski algoritam s više razlučivosti za značajno poboljšanje performansi postojećih diskretnih optimizacijskih algoritama. Problem CPO se zatim preformulira u domeni kontinuiranog prostora i uspoređuje s varijantom diskretnog prostora kako bi se istaknula poboljšana točnost. Na kraju, metoda segmentacije supervoksela, koja je skrojena za upotrebu u metodi optimizacije više razlučivosti, uvedena je i potvrđena na dobro poznatim skupovima podataka 3D oblaka točaka. Eksperimenti i rezultati simulacije na 3D modelima visoke razlučivosti raznih vozila pokazuju da su predložene metode učinkovite u optimiziranju položaja više kamera za okružni prikaz vozila, ispunjavajući zahtjeve scenarija iz stvarnog svijeta u razumnom vremenu.
Keywords
camera placement optimization
global optimization
image segmentation
computational geometry
3D visualization
Keywords (croatian)
optimizacija položaja kamere
globalna optimizacija
segmentacija slike
računalna geometrija
3D vizualizacija
Language english URN:NBN urn:nbn:hr:168:210870 Promotion 2024 Study programme Title: Computer Science Type of resource Text Extent 97 str. File origin Born digital Access conditions Open access Terms of use Created on 2024-08-27 07:39:21
