Abstract
This paper describes a machine learning approach for visual pattern recognition framework that is capable of processing images rapidly while achieving high recognition rates. The key contribution of this study is a tractable model suitable for object recognition, that takes advantage of two complementary components: (a) a bag of covariance matrices object descriptor based on feature selection, allowing to combine the advantages of histogram and appearance model with the ability of scalability, (b) a model with incremental learning based on on-line variant of Random Forests (RF) learning algorithm. We evaluate the potential of the proposed procedure with empirical studies in the domain of the GRAZ02 dataset. The system yields in object recognition performance comparable to the benchmark standard RF, AdaBoost, and SVM classifiers. Experimental results demonstrate significant performance improvement.
