On statistical approaches to target silhouette classification in difficult conditions

In this paper we present a methodical evaluation of the performance of a new and two traditional approaches to automatic target recognition (ATR) based on silhouette representation of objects. Performance is evaluated under the simulated conditions of imperfect localization by a region of interest (ROI) algorithm (resulting in clipping and scale changes) as well as occlusions by other silhouettes, noise and out-of-plane rotations. The two traditional approaches are holistic in nature and are based on moment invariants and principal component analysis (PCA), while the proposed approach is based on local features (object parts) and is comprised of a block-by-block 2D Hadamard transform (HT) coupled with a Gaussian mixture model (GMM) classifier. Experiments show that the proposed approach has good robustness to clipping and, to a lesser extent, robustness to scale changes. The moment invariants based approach achieves poor performance in advantageous conditions and is easily affected by clipping and occlusions. The PCA based approach is highly affected by scale changes and clipping, while being relatively robust to occlusions and noise. Furthermore, we show that the performance of a silhouette recognition system subject to mismatches between training and test angles of silhouettes (caused by an out-of-plane rotation) can be considerably improved by extending the training set using only a few angles which are widely spaced apart. The improvement comes without affecting the performance at “side-on” views.