Covariance-based recognition using an incremental learning approach

We propose an incremental machine-learning approach for object recognition where new images are continuously added and the recognition decision is made with no delay. First, the object region is automatically represented using a bag of covariance features. Then an on-line variant of the random forest (RF) classifier is employed to select object descriptors and to learn the object classifiers. A validation of the method by empirical studies in the domain of the GRAZ02 dataset shows its superior performance over those methods which are histogram-based, and subsequently yields in object recognition performance comparable to that of state-of-the-art classifiers.