基于方向描述符的物体检测

Object Detection Based on Orientation Descriptor

对于形状和表面纹理都有变化的物体的检测,局部不变性算子并不适用,而现有的局部描述符对于区分这种形状的作用也并不明显。为此本文提出了一种新的基于方向描述符的物体检测算法。根据模型轮廓图或边缘图像计算出初始描述符,在此基础上为图像中的每一点生成方向描述符。方向描述符既可以描述边界的走向,又可以容忍边界的较小变形。使用多分辨率加速的滑动窗口算法,将每个有效的候选区域与模型的描述符矩阵进行匹配,以判断此位置是否包含目标物体。实验结果显示,本文算法取得了相对较高的检测率。

Local invariant algorithms are usually not applicable for object detection with variance of shape and surface texture, and existing local descriptors show limited ability to distinguish this kind of shape. A new object detection algorithm based on orientation descriptor is proposed. Initial descriptors are calculated based on silhouette of model or edge image of testing image, on this basis of which, orientation descriptor are calculated for each pixel in image. Orientation descriptor can describe edge orientation and tolerate small shape deformation. Multi-resolution method is utilized to speed up sliding-window algorithm. Descriptor matrixes of valid candidate region and model are matched to determine the existence of object at this position. The experimental results show that the proposed algorithm achieves fine detection rate relatively.