Research and Development (1987-1992)

Development of three-dimensional thinning algorithms 

Based on a popular 2-D thinning algorithm (SPTA) two new algorithms for 3-D thinning (called as ESPTAand MESPTA) have been designed with satisfactory performance. 

Development of segmentation algorithms for 3-D scene analysis 

Concepts of Digital Neighbourhood Plane and Neighbourhood Plane Set (NPS) have been introduced for segmenting three-dimensional scenes. Using these a novel region growing algorithm has been developed for both voxel data and range data. The NPS provides a very simple yet sufficiently useful information of the orientation of surface points. The relationships between the NPS values of the segments are used to obtain the structural information of the scene easily. NPS is also the quantitative measure used in the qualitative object description system and is easy to compute and use in 3-D algorithms. Its use greatly increases both the speed and the flexibility of the system. 

Development of algorithms for wire frame representation of 3-D objects 

Wire framing technique for the 3-D binary objects has been developed. This is robust with respect to object rotation and tolerant to small distortions. This makes the wire frame graph a viable scheme for object modeling. Wire framing can also be used for modeling surfaces and graphics display of an unknown object shape. The technique has also been extended to range images.  

Development of methodologies for higher level scene understanding 

Using the lower level information higher level scene understanding in terms of recognition of three dimensional objects has been achieved. Two different situations have been considered in this regard. In the first case recognition of single isolated three dimensional object in the form of a 3-D binary array has been carried out. The 3-D object is represented by a wire frame graph using the algorithm mentioned above and then a fuzzy graph matching has been carried out with the specified wire frame models of different object class. Different new similarity measures using t-cost metrics are used for this purpose.

In the second scheme a hierarchical qualitative description of a 3-D multiple object scene (in the form of a range image) is obtained. In this scheme it is shown that in a recognition system ,which accomodates a wide class of objects, qualitative reasoning for a broad object class description before applying detailed mathematical techniques for classification at the initial stage of the recognition system, enhances the speed of recognition.  

Development of parallel vision algorithms 

Parallel algorithms for range image segmentation in a Pyramid Machine has been developed. A connectionist methodology for detecting peaks of a distribution over a multidimesional space has also been developed. This is subsequently used in the segmentation of intensity images , detection of straight lines in intensity images , segmentation of colored images and multi-spectral satellite images .  

Analysis of dynamic stereo images 

Different preprocessing algorithms such as tracking , segmentation, corner detection, detection of occlusion etc. has been developed. These are used for 3-D motion estimation from dynamic stereo images.