Learning-based 3D Geometric Modeling For Concept Design Of Vehicle Body And Its Applications

Automobile styling design is a process from 2D sketches to 3D models.Designers are responsible for showing their creative ideas by drawing different sketches on the paper and the 3D stylists are responsible for converting 2D hand drawn sketches to 3D digital models using CAD technology which is short for Computer Aided Design.The operation of current commercial CAD software is so complex that it is impossible to convert your own hand drawn vehicle sketches to corresponding 3D models rapidly and automatically and realize "what you see is what you get".So in this paper,the automatic understanding of 2D vehicle side view is discussed and based on this,3D digital model is reconstructed efficiently.This fully automated modeling process will significantly improve the design efficiency of vehicle body model and lower design threshold.It is of great significance to the innovative design of automobile.This paper presented a data-driven and machine learning modeling approach to realize the fully automated 3D modeling of vehicle body based on single view(side view)renderings or freehand sketching.Firstly,vehicle model recognition and classification is realized based on the input image.And then,landmarks of vehicle side view are localized by regressing local binary features.finally,2D and 3D curves of vehicle body are reconstructed according to the database of 2D and 3D vehicle body form feature lines.The reconstructed 2D and 3D curves can be used to update the database for further reconstruction.Here are the main works of this paper.(1)The creation of annotated vehicle side view database and the realization of vehicle model recognition and classification based on vehicle side view.By analyzing the common passenger vehicles with four doors,we divided vehicles into two types,sedan and SUV.Then according to the location of back quarter glass,we divided each type into another two types,vehicle with back quarter glass and vehicle without back quarter glass.We created annotated vehicle side view database which is composed of 4157 images and presented a combined vehicle model recognition and classification approach.Vehicle model recognition and classification is the foundation of landmark localization and 2D and 3D vehicle curve and surface reconstruction.(2)The creation of vehicle side view landmarks database and the recognition and localization of vehicle side view landmarks by regressing local binary features.We defined 32 points as landmarks by analyzing the structure of vehicle side view and reconstruction error rate and drawing lessons from facial landmarks definition.We created vehicle side view landmarks database which is composed of 4157 images.The landmarks localization result also is the foundation of 2D and 3D vehicle curve and surface reconstruction.(3)The reconstruction of 2D and 3D vehicle curves and surface and the update of its database.We applied the characteristic curve network with consistent representation to represent 3D vehicle body model.We realized the 2D and 3D vehicle curves reconstruction based on statistical morphable model.These reconstructed 2D and 3D curves can update the database if needed so that this data-driven reconstruction approach can be more precise.This modeling approach mentioned above can reconstruct 3D vehicle model from vehicle side view automatically and it is proved to realize the automatic reconstruction of vehicle concept model based on side view efficiently and precisely through a series of numerical experiments.