DURIP: Object Recognition in Cluttered Scenes Using Compressed Data

摘要

The PI requested funds for the acquisition of a 3D sensor for capturing range data and for high performance computing workstations to renovate Penn State's Robust Purposive Vision Laboratory. This equipment is supporting research towards the development of both theoretical tools and specific algorithms for robust object recognition using compressed data. This research was initiated by the PI during her 1997 summer research visit to Eglin Air Force Base and it is the subject of a proposal to be submitted to AFOSR in the near future. Robust, reliable methods for automatic target/object detection, recognition, classification and identification are key technology areas for meeting the U.S. Air Force requirements for defense operations in warfare, as well as in peacekeeping and humanitarian role situations. However, the recognition of general three-dimensional objects in cluttered scenes remains a challenging problem. In particular, the design of a good representation suitable to model large numbers of generic objects that is also efficient and robust to occlusion and segmentation problems, while minimizing probabilities of false alarm and misdetection, has been an stumbling block in achieving success. To address this problem we have developed a new representation and theoretical models for object recognition based on appearance-based parts (ABPs) and relationships (ABRs), obtained from collections of images compressed using the Wavelet transform. This representation will allow us to design a recognition system that overcomes the problems mentioned above and that can work directly on compressed data, during both the training and the recognition stages, making it both time and memory efficient.