基于HOG-LBP特征的自适应实时手掌检测算法

HOG-LBP Based Feature Adaptive Real-time Palm Detecting Algorithm

论文提出了一种将HOG特征和LBP特征相融合的新方法,用来识别无限制的彩色图像中的手掌。融合后的特征能更好地描述手掌的边沿特征和子结构,并且保持了对光照和复杂背景的不敏感性。由于手掌的宏观结构特征含有最多的可用于判别的信息量,所以可以将HOG特征和LBP特征相融合来处理手掌检测。最后,将HOG和LBP融合后的特征用线性SVM分类进行分类。实验结果表明,在该文建立的手掌数据库上,性能得到了明显的提升。

This paper studied the HOG feature and LBP features, and integrated them together which will result in the increase of the accuracy in feature detection. Here, we named this method adaptive HC^LBP feature fusion algorithm. The integrated features will be clas- sified by SVM. Since HOG features works within the scale of a local cell unit, it can avoid the optical and geometric image deformation, which can be well applied to the detection of an object. Furthermore, the advantage of LBP is the speed of calculation on light robustness, but due to LBP window size is fixed and has nothing to do with the image, LBP texture primitives in the feature extraction errors, it is diffi- cult to adapt to different scales roughness and texture requirements. There are two traditional fusion methods, one is the co-training. This approach generally needs to use two or more classifiers to accomplish the goals of recognition, and can achieve better tracking performance. However, based on the multiple classifiers, there exits the requirement of complex calculations. The other method is to directly merge fea- ture matrix. This method is simple, but not well characterized by the combination of multiple features. In this paper, based on the advanta- ges and disadvantages of the above two methods, the writer proposed an adaptive HOG-LBP fusion algorithm, which can be a good fusion of HOG and LBP features, which will not only guarantee the performance but reduce the amount of computation.Experimental results show that the method can identify and track the palm, and adapt to low-quality color images on a PC platform to achieve real-time processing per- formance.