TY - JOUR
T1 - Detección de Automóviles en Escenarios Urbanos Escaneados por un Lidar
AU - Ramírez-Pedraza, Alfonso
AU - González-Barbosa, José Joel
AU - Ornelas-Rodríguez, Francisco Javier
AU - García-Moreno, Angel Iván
AU - Salazar-Garibay, Adan
AU - González-Barbosa, Erick Alejandro
N1 - Publisher Copyright:
© 2014 CEA. Publicado por Elsevier España, S.L.U. Todos los derechos reservados.
PY - 2015
Y1 - 2015
N2 - Detection of vehicles on 3D point clouds is performed by using the algorithm presented in this work. Point clouds correspond to urban environments and were acquired with the LIDAR Velodyne HDL-64E. The environment is considered semi-structured so that can be modeled using planes. Vehicle detection is carried out on to stages, segmentation and indexation. First stage is at the same time composed of three sub-stages. In the first one the principal plane (in this case the floor) is extracted, in the second sub-stage secondary planes are extracted using a tailored version of Hough's method, secondary planes are those perpendicular to the main plane. Finally in the third sub-stage and using MeanShift method, the remaining objects are segmented. Indexation on its side is divided into two sub-stages, in the first one, last segmented objects using MeanShift method are modeled using histograms according to the direction of the object's 3D points normal; in the second stage histograms are compared to those previously stored on a database of object's histograms. Optimizing of detection thresholds was carried out through ROC analysis. Two databases were used during the experiments, the first DB have 4500 objects and was used for ROC analysis training; the second one contained 3000 objects and was used for verification.
AB - Detection of vehicles on 3D point clouds is performed by using the algorithm presented in this work. Point clouds correspond to urban environments and were acquired with the LIDAR Velodyne HDL-64E. The environment is considered semi-structured so that can be modeled using planes. Vehicle detection is carried out on to stages, segmentation and indexation. First stage is at the same time composed of three sub-stages. In the first one the principal plane (in this case the floor) is extracted, in the second sub-stage secondary planes are extracted using a tailored version of Hough's method, secondary planes are those perpendicular to the main plane. Finally in the third sub-stage and using MeanShift method, the remaining objects are segmented. Indexation on its side is divided into two sub-stages, in the first one, last segmented objects using MeanShift method are modeled using histograms according to the direction of the object's 3D points normal; in the second stage histograms are compared to those previously stored on a database of object's histograms. Optimizing of detection thresholds was carried out through ROC analysis. Two databases were used during the experiments, the first DB have 4500 objects and was used for ROC analysis training; the second one contained 3000 objects and was used for verification.
KW - 3D Segmentation
KW - 3D point cloud
KW - Lidar
UR - http://www.scopus.com/inward/record.url?scp=84941953797&partnerID=8YFLogxK
U2 - 10.1016/j.riai.2015.03.003
DO - 10.1016/j.riai.2015.03.003
M3 - Artículo
SN - 1697-7912
VL - 12
SP - 189
EP - 198
JO - RIAI - Revista Iberoamericana de Automatica e Informatica Industrial
JF - RIAI - Revista Iberoamericana de Automatica e Informatica Industrial
IS - 2
ER -