On classifying silhouettes in adverse conditions

We compare the performance of holistic and local feature approaches for the purpose of classifying silhouettes in adverse conditions (i.e. occlusions by other silhouettes, noise and imperfect localization by a region of interest algorithm, resulting in clipping and scale changes). Holistic feature extractors based on Hu's moment invariants and principal component analysis (PCA) are coupled with a classifier based on Gaussian densities, while a local feature extractor based on the 2D Hadamard transform (HT) is coupled with a Gaussian mixture model (GMM) based classifier. Experiments show that the HT/GMM approach is relatively robust to clipping, scale changes and occlusions; however in its current form it is highly sensitive to noise. The results further show that the moment based approach achieves relatively 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.