Steps of the work 1 – Gaussian Mixture Models Implemented in C/C++ based in Cluster Library from Charles A. Bouman (Purdue University) Automatically compute the number of GMM Speedup but it isnt realtime with a lot of points (>5000) yet. Look for a method to reduce the number of points (size of the grid in occupancy grid) Generate more than one GMM to one “shape”, solve it with a prior knowledge.
Steps of the work 2 – Change Detection Implemented in C/C++ using EMD library from Carlo Tomasi (Duke University) Works fine using a greedy algorithm (Fast) Possible implements something to detect false positive
Steps of the work 3- Shape Retrievel Implemented in C/C++ based in the library levmar from Manolis Lourakis (Foundation for Research and Technology – Hellas) Hard to solve underdetermined problem, 6 variables but only 3 equation(using 3 eigenvalues). Need a bayes classifier ? Problem to compute the eigenvalue of non- symmetric matrix (generalized eigenvalue) – Look for a new library
Steps of the work 3- Shape Retrievel Instead of using basic shapes(plane,sphere), using a superquadrics less than 1: a pointy octahedron with concave faces and sharp edges. exactly 1: a regular octahedron. between 1 and 2: an octahedron with convex faces, blunt edges and blunt corners. exactly 2: a sphere greater than 2: a cube with rounded edges and corners. infinite (in the limit): a cube
Steps of the work 4- Integration with Carmen/Tests Meeting with João Monteiro to define details. Start with 2D maps and Change Detection After implemented the 3D maps, try put the shape retrievel in Carmen.
Some tests Real Data Old Data – The Door (2160 points) New Data – The Corridor with the door (40000 points) Uknown the Number Classer – 20 Class the maximun size Time Processing – 118 s
Some tests 2- Simulated data Old Data – Corridor (1323 points) New Data - Corridor + Desk + Ball (2960 points) Time processing – without knowledge 3.8 s / with knowledge 3 s