Objektituvastus maastikul kasutades lidarit ja kaamerat
Files
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Seoses hüppelise huvi kasvuga autonoomsete sõidukite vastu viimastel aastatel on suurenenud ka vajadus täpsemate ja töökindlamate objektituvastuse meetodite järele. Kuigi tänu konvolutsioonilistele närvivõrkudele on palju edu saavutatud 2D objektituvastuses, siis võrreldavate tulemuste saavutamine 3D maailmas on seni jäänud unistuseks. Põhjuseks on mitmesugused probleemid eri modaalsusega sensorite andmevoogude ühitamisel, samuti on 3D maailmas märgendatud andmestike loomine aeganõudvam ja kallim. Sõltumata sellest, kas kasutame objektide kauguse hindamiseks stereo kaamerat või lidarit, kaasnevad andmevoogude ühitamisega ajastusprobleemid, mis raskendavad selliste lahenduste kasutamist reaalajas. Lisaks on enamus olemasolevaid lahendusi eelkõige välja töötatud ja testitud linnakeskkonnas liikumiseks.Töös pakutakse välja meetod 3D objektituvastuseks, mis põhineb 2D objektituvastuse tulemuste (objekte ümbritsevad kastid või segmenteerimise maskid) projitseerimisel 3D punktipilve ning saadud punktipilve filtreerimisel klasterdamismeetoditega. Tulemusi võrreldakse lihtsa termokaamera piltide filtreerimisel põhineva lahendusega. Täiendavalt viiakse läbi põhjalikud eksperimendid parimate algoritmi parameetrite leidmiseks objektituvastuseks maastikul, saavutamaks suurimat võimalikku täpsust reaalajas.
Since the boom in the industry of autonomous vehicles, the need for preciseenvironment perception and robust object detection methods has grown. While we are making progress with state-of-the-art in 2D object detection with approaches such as convolutional neural networks, the challenge remains in efficiently achieving the same level of performance in 3D. The reasons for this include limitations of fusing multi-modal data and the cost of labelling different modalities for training such networks. Whether we use a stereo camera to perceive scene’s ranging information or use time of flight ranging sensors such as LiDAR, the existing pipelines for object detection in point clouds have certain bottlenecks and latency issues which tend to affect the accuracy of detection in real time speed. Moreover, these existing methods are primarily implemented and tested over urban cityscapes.This thesis presents a fusion based approach for detecting objects in 3D by projecting the proposed 2D regions of interest (object’s bounding boxes) or masks (semantically segmented images) to point clouds and applies outlier filtering techniques to filter out target object points in projected regions of interest. Additionally, we compare it with human detection using thermal image thresholding and filtering. Lastly, we performed rigorous benchmarks over the off-road environments to identify potential bottlenecks and to find a combination of pipeline parameters that can maximize the accuracy and performance of real-time object detection in 3D point clouds.
Since the boom in the industry of autonomous vehicles, the need for preciseenvironment perception and robust object detection methods has grown. While we are making progress with state-of-the-art in 2D object detection with approaches such as convolutional neural networks, the challenge remains in efficiently achieving the same level of performance in 3D. The reasons for this include limitations of fusing multi-modal data and the cost of labelling different modalities for training such networks. Whether we use a stereo camera to perceive scene’s ranging information or use time of flight ranging sensors such as LiDAR, the existing pipelines for object detection in point clouds have certain bottlenecks and latency issues which tend to affect the accuracy of detection in real time speed. Moreover, these existing methods are primarily implemented and tested over urban cityscapes.This thesis presents a fusion based approach for detecting objects in 3D by projecting the proposed 2D regions of interest (object’s bounding boxes) or masks (semantically segmented images) to point clouds and applies outlier filtering techniques to filter out target object points in projected regions of interest. Additionally, we compare it with human detection using thermal image thresholding and filtering. Lastly, we performed rigorous benchmarks over the off-road environments to identify potential bottlenecks and to find a combination of pipeline parameters that can maximize the accuracy and performance of real-time object detection in 3D point clouds.