Integral Channel Feature based arbitrary object tracking

Object tracking is a challenging problem in computer vision as many performance affecting factors need to be considered in a robust algorithm. We propose a framework to consolidate Integral Channel Features (ICF) to represent targets' appearance by embedding global and patch based approaches which offer feature strength and accuracy to the target template. The use of ICF expedites the extraction of color and structural features from video frames in a very efficient manner. Application of the patch based approach on global templates with maximum similarity metric enables better object representation. Target's appearance representation is updated using k-means clustering and vector quantization. We use incremental PCA learning for acquiring training samples and presenting fixed size feature codebook vectors. Experiments are conducted to compare performance between the proposed approach and two other state of the art approaches. Results show that the proposed approach outperforms published state of the art methods.