A Novel Feature for Object Detection

This paper proposed a new 2-dimension-code-like feature based on the Haar-like feature proposed by Viola et al. The feature can be calculated in different scales rapidly once the integral image is calculated and this characteristic is inherited from the Haar-like feature. Instead of reducing the dimensions after calculating redundant features, we calculate feature vectors with fixed dimensions for each sub-window, so the training speed is greatly improved. To test our feature, we build a face detector based on AdaBoost. The results of the face detector show that this feature can take full advantage of the image spatial location information and also insensitive to illumination changes.