Deep learning for object tracking in 3D point clouds

View/Open
Cita com:
hdl:2117/329126
CovenanteeTechnische Universität München
Document typeBachelor thesis
Date2020-06
Rights accessOpen Access
All rights reserved. This work is protected by the corresponding intellectual and industrial
property rights. Without prejudice to any existing legal exemptions, reproduction, distribution, public
communication or transformation of this work are prohibited without permission of the copyright holder
Abstract
Great progress has been achieved in computer vision tasks within image and video; however, technological advances in LiDAR sensors have created a whole new area of computer vision research devoted to it. With applications in many industries, such as transportation, agriculture, or healthcare. This thesis studies object tracking in 3D point clouds. Pairs of point cloud observations are feed to a neural network to estimate pose and translation between the observations. Then these estimations, together with external features, are processed with Kalman Filter and RNN to extract spatial-temporal redundancies and improve the results. The system has been tested in the KITTI dataset, with pre-segmented observations, on different types of objects and paths. The results show that the neural network estimated pose gives a very accurate tracking. Still, the best results are achieved when combining the estimated pose and translations with a recurrent neural network. Great progress has been achieved in computer vision tasks within image and video; however, technological advances in LiDAR sensors have created a whole new area of computer vision research devoted to it. With applications in many industries, such as transportation, agriculture, or healthcare. This thesis studies object tracking in 3D point clouds. Pairs of point cloud observations are feed to a neural network to estimate pose and translation between the observations. Then these estimations, together with external features, are processed with Kalman Filter and RNN to extract spatial-temporal redundancies and improve the results. The system has been tested in the KITTI dataset, with pre-segmented observations, on different types of objects and paths. The results show that the neural network estimated pose gives a very accurate tracking. Still, the best results are achieved when combining the estimated pose and translations with a recurrent neural network. Great progress has been achieved in computer vision tasks within image and video; however, technological advances in LiDAR sensors have created a whole new area of computer vision research devoted to it. With applications in many industries, such as transportation, agriculture, or healthcare. This thesis studies object tracking in 3D point clouds. Pairs of point cloud observations are feed to a neural network to estimate pose and translation between the observations. Then these estimations, together with external features, are processed with Kalman Filter and RNN to extract spatial-temporal redundancies and improve the results. The system has been tested in the KITTI dataset, with pre-segmented observations, on different types of objects and paths. The results show that the neural network estimated pose gives a very accurate tracking. Still, the best results are achieved when combining the estimated pose and translations with a recurrent neural network.
SubjectsMachine learning, Artificial intelligence, Neural networks (Computer science), Robot vision, Aprenentatge automàtic, Intel·ligència artificial, Xarxes neuronals (Informàtica), Visió artificial (Robòtica)
DegreeGRAU EN ENGINYERIA DE TECNOLOGIES I SERVEIS DE TELECOMUNICACIÓ (Pla 2015)
Files | Description | Size | Format | View |
---|---|---|---|---|
thesis.pdf | 2,826Mb | View/Open |