基于协同重排序的手势识别方法

The pose estimation of hand is a theoretically interesting and challenging problem in computer vision with many applications such as human-computer interaction, sign language recognition, virtual character control and so on. However, due to the high degree of freedom (DoF) of the hand pose, it is very difficult, if not impossible, to directly estimate the hand pose efficiently. In this paper, we proposed a hand pose estimation method, namely, collaborative reranking. It divides the observation of hand into many partial observations, which is related with a subset of the phalanges joint angles. And it builds a partial observation database for each partial estimator position off-line. At pose estimation stage, it extracts partial observations from the depth image, and estimate the parameters for each partial observation by k nearest neighbors (k-NN) searching. Then it reranks the k-NN of each partial observation according to the k-NN searching result of its neighbor partial estimators. We model this idea into a graph model, and obtain the collaborative reranking algorithm by systematically and rigorously mathematical inference. Although collaborative reranking mainly focuses on hand local pose estimation, we also proposed a method to estimate the hand global motion to make the system usable. Finally, we verify the performance of the proposed method by experimenting on synthetic and realistic depth image. The proposed method can estimate hand pose within 30 ms (17 ms for local pose estimation and 12 ms for global pose estimation) without GPU speedup, and the maximum average estimation error is less than 10°. Extensive experimental results demonstrated the efficiency and effectiveness of the proposed method.