Integrated range image segmentation using connectionist paradigm

Flexible robotic and automation systems must have the capability to sense and interpret (or analyze) their surroundings to enable them to react accordingly. Such automated systems are desirable for a number of manufacturing tasks such as assembly, material handling and transportation, inspection, obstacle avoidance, etc. Range sensors are becoming increasingly popular in automated systems for gathering 3-dimensional (3-D) information about the surroundings. This paper presents a two-stage connectionist neural net model which extracts surface features and groups them via local interactions. The first stage optimally computes surface parameters, e.g., surface normals, curvature and discontinuities (crease or jump). In the second stage, surface features compete with each other to group similar image pixels. Daugman´s projection net (DPNN) and Kohonen´s self organizing neural net (KSNN) are used for feature extraction and region growing, respectively. Experimental results are included to demonstrate the performance of the proposed technique