多重空间特征融合的手势识别

Hand Gesture Recognition Using Multiple Spatial Features Fusion

手势识别是计算机视觉领域的一个重要课题,有着广泛的应用,如交互式游戏和手语识别等.随着深度传感器的面世,手势识别任务变得更为简单.近些年有大量的方法尝试在深度图像中提取特征,来作为某种手势的有效表达.但由于手势固有的灵活性和复杂性,现有算法在大型数据集上的识别效果依然不能令人满意.本文提出一种新的基于多重空间特征融合的方法来识别静态的手势深度图像,即对三维点云进行局部的主成分分析,并提取局部的梯度信息和局部点云的深度分布,这些信息有效的编码了手势的局部形状,本文把局部特征连接起来作为整个手势图像的特征,并通过随机森林分类器的分类结果对特征进行过滤,从而剔除对分类结果没有影响的特征.最后用过滤后的特征再次训练随机森林来识别手势.与当下流行的手势识别算法相比,本文的方法在两个大型手势数据集上有效的提高了识别率.

Hand gesture recognition is an important topic of computer vision. It has a wide range of applications,such as sign language recognition and interactive computer games. With the launch of the depth sensors,hand gesture recognition becomes an easier task than before. Large amounts of methods have attempted to extract features from depth image,as a valid expression of certain kind of gesture.However,due to the inherent flexibility and complexity of human hand,existing algorithms still perform poorly on large datasets. This paper presents a novel approach to classify static hand postures based on the integration of multiple spatial features from depth images.We perform PCA on the local patch of 3D cloud points,combined with local gradient information and local depth distribution. These information effectively encoded local shape of hand. We concatenate local features to form a global descriptor. Then we consider feature pruning from outputs of randomized decision forest to delete those are not significant for classification. Finally,we train RDF again with discriminative features to classify hand gestures. Compared with the state-of-the-art methods,our methods effectively improve the recognition results on two large hand gesture datasets.