Realtime Coarse Pose Recognition Using a Multi-scaled Local Integral Histograms

We present a fast and robust algorithm for segmenting foreground object from background image by comparing local histograms. Background subtraction is a important preprocessing step for extracting the features that can be used for object tracking in surveillance system or HCI system in virtual environment. In this paper, the local histograms of the same area are used to compute a foreground probability. The histogram-based method is partially robust against illumination change and small moving objects in background. However without data quantization to reduce bin size, histograms are generally not suitable for realtime applications. Moreover quantization errors are a major drawback of using histograms. We propose a new method to keep the advantages of histograms without suffering computational load and quantization error using local kernel histogram with the multi-scaled integral histograms. We implement the video game interface with a trained neural network to prove the proposed method is highly applicable to coarse pose recognition.