Computer Vision Technology
New Computer Vision Technology for Object Identification on Patterns and 3D shape Reconstruction on Stereo Image.
Tech Area / Field
- FIR-NSS/Nuclear Safety and Safeguarding/Fission Reactors
- SAT-VEC/Space Launch Vehicles and Support Equipment/Space, Aircraft and Surface Transportation
- MAN-ROB/Robotics/Manufacturing Technology
3 Approved without Funding
State Institute of Physics and Technology, Russia, Moscow
- Academician Pilugin Scientific-Production Center of Automatics and Instrument-Making, Russia, Moscow
Project summaryThe project purposes:
The project provides for a development of three computer vision methodologies. The first methodology is intended for automatic object identification based on fast pattern recognition, the second one is to measure 3-D coordinates of scene points chosen by an operator on one of two or more stereo images, and the third methodology ensures 3-D surface reconstruction with a time series of images processed in real time. The object identification methodology uses the tree-like invariant structures to represent the patterns of the objects and a dictionary code tree of the model patterns. These data structures allow essentially to speed up the recognition procedure relative to the known methods. The stereo non-contact measuring methodology is based on using two, three or more stereo images of the same 3-D scene. The measuring procedure is universal to the number of images. This property differs the offered methodology from the known ones and allows to increase a precision and a reliability of the measuring procedure as a number of images increases. Both object identification and stereo measuring methodologies can be used to create computer vision systems in industrial and medical applications and for robots.
The technical approaches:
Development of the object identification methodology based on pattern recognition methods includes the following elaboration:
- moment-based representation of patterns with hierarchically structured geometric primitives;
- creation of a dictionary code tree of the representations for a set of model patterns using coding and sorting methods;
- elaboration of a fast search algorithm in the dictionary tree for a model with the most similarity to a given object.
Development of the methodology of measuring 3-D point coordinates based on stereo images includes the following elaboration:
- representation of two or more stereo images by their features in appropriate attribute space;
- identification of the features of stereo images using their representations;
- calculation of 3-D coordinates for selected points identified on stereo images.
Development of 3-D surface reconstruction methodology using time series of images includes the following elaboration:
- determination of parameters and errors of multispectrum sensor and inertial navigation system;
- evaluation of reconstruction errors by means of simulation and statistical estimation;
- identification of points on time series of images;
- elaboration of 3-D reconstruction algorithms.
Expected Results. Due to activities carrying out within the project the following results are expected. The advanced approaches will be developed for the object identification by means of moment-based fast pattern recognition method and for non-contact measuring 3D coordinates with two, three or more stereo images. It is planned to obtain estimations of a precision of measuring and a complexity of processing. Elaboration of the object identification and the stereo measuring methodologies supposes to create a laboratory model of computer vision system and appropriate demonstration programs.
In a frame of 3D reconstruction methodology the limitations (restrictions) of 3-D surface reconstruction caused by different errors will be revealed. The comparative estimations of an effeciency of different algorithms for setting up the correspondent points in time series of images will be obtained. Efficient indexes of different algorithms will be supplemented by an estimation of computational complexity and an estimation of weight and sizes of a special processor executing 3-D shape reconstruction in real time. It is planned to elaborate algorithms and demo programs that realise 3-D surface reconstruction methodology.
Potential role for Foreign collaborators:
- exchange of information during project implementation;
- joint seminars, workshops, meetings, consultations;
- verification of results using independent methods and or equipment.
The methodology of 3-D surface reconstruction is intended creating a computer vision system to be applied for controlling trajectories of air and spacecrafts. Such control systems have more high precision and stability relative to the current disturbances of both artificial and natural origin. The characteristics are achieved with a correction of inertial navigation system errors by means of matching a viewed scene and a digital map or by recognition of preliminary determined objects. In addition to the navigation systems, 3-D surface reconstruction can be used in automatic landing systems of both aircraft and spacecraft.