Abstract:This paper presents an improved SIFT(Scale Invariant Feature Transform) descriptor for local feature detection and matching in object tracking. Only the local maxima in DOG scale space are detected as candidate interesting points to improve the stability. In order to avoid rotating the image, the main orientations and descriptors are determined statistically, according to oriented gradients histograms in circular neighborhood around the interesting point. Finally, ratio between the first and the second closest distance is used to match the 96-dimensional vectors. This method exhibits very good performance in high reliable applications, for its effectiveness and reduced complexity.