Real Time Object Localization and Recognition from Silhouette Images

摘要

This paper describes a new method for real time online 2D detection and recognition of objects from silhouette images based on learning of multiple projections. Compared to recognition methods based on geometric moments, Fourier descriptors or Blob analysis this approach has several important advantages. It can deal with multiple objects, which are touching. The method is able to detect both global as well as small local differences between objects. The algorithm is trained with only one sample image. After training it can detect and recognize this object among other objects and retrieves position and orientation. The algorithm tolerates object rotation up to a range, which can be specified by the user during training. A hierarchical processing sequence, with coarse search in 1D projections and fine search in 2D images results in a very fast inspection procedure. The algorithm is able to analyze one image (320x240 pixels) within 5 ms on a PC (Pentium III 550Mhz).