Real-Time Tracking with Selective DoP-RIEF Features for Augmented Reality
Yi Zhang,Ping Lu,Jie Chen,Lingyu Duan
DOI: https://doi.org/10.1109/BigMM.2015.30
2015-01-01
Abstract:Real-time, accurate and robust target tracking on mobile devices is an important problem which can facilitate applications such as augmented reality. However, it is still unsolved, partly due to the mobile's computing limitations. Compressive tracker performs favorably against state-of-the-art algorithms in terms of efficiency, accuracy and robustness, but as limited by the speed of feature matching, it cannot achieve real-time tracking in mobile applications. In this paper, we propose a fast feature, i.e., selective Difference of Patch Robust Independent Elementary Features (DoP-RIEF). DoP-RIEF is a global feature which is related to BRIEF. It uses histogram to fit feature distribution because it is more flexible than Gaussian, and intermediate results for subsequent classification can be stored, avoiding duplication of operations. Feature selection further deletes features which are less discriminative and improves the feature quality. Through these two steps, the feature matching can be accelerated significantly and at the same time tracking accuracy and robustness are improved. Compared with compressive tracker on 17 publicly available sequences, our method outperforms it in terms of both robustness and accuracy. In addition, the speed is about 270 frames per second which is 8 times faster than the compressive tracker. To further evaluate our algorithm in natural scenes with obvious scale, rotation, and illumination variations, we test it on Stanford datasets and Peking University landmark datasets, and the accuracy is above 90%.