A bank of classifiers for robust object modeling in wavelet domain

Image and video content analysis applications typically require functionalities such as object classification, detection and tracking, and activity recognition. Objects may undergo translation, rotation, and changes in scale due to perspective projection. Further, the appearance of objects and illumination conditions may change over time. Occasionally objects might also occlude one another in the scene making consistent classification, detection, and tracking a challenge. To reduce the effect of these limitations it is proposed to model objects in wavelets domain using silhouettes. The silhouette of an object is characterized using projected histograms of sixteen wavelet primitives extracted from a silhouette map of the scene. A classifier based on eigen decomposition of histogram of feature vectors combined with sparse coding prediction is presented. The model of a class is represented as over complete dictionary of sparse codes. For robustness multiple classifiers based on the same sparse code operate in parallel but at different scales. It is combined with spatial histogram classifier to realize a bank of multiple classifiers. The accuracy of the proposed classifier is compared with support vector machine and published state-of the-art results. Accuracy evaluation and real-time performance demonstrates competitive performance with the published stat-of-the-art results.