Object detection based on image/video in the scene, has always been to involve how to deal with the geometry and scale deformation, as well as the changes in the form of movement caused by imaging perspective, which makes the algorithm design is particularly complex in 2D image space, especially for object shelter and adhesion problems, always can not be well resolved.Considering of this, a new target detection method based on mixing component models in 3D space is proposed.Firstly, parallel to the surface of target, appropriate inverse projective planes or array on 3D space are established.Then, through the inverse projection transformation, the 2D image is projected inversely to the real 3D space and the inverse projection images of target apparent characteristics are formed.After that, component dictionaries are learned based on the HOG feature of target in the inverse projection images, and on account of this, sparse decomposition approach is adopted to implement the detection of local components.Finally, for the candidate components, centroid clustering in a 3D model is used to further identify the target.Experiment results indicate that the object detection method based on 3D multi-components model and inverse projection in image/video not only have low complexity, but also can perfectly deal with the object shelter and adhesion problem, and the accuracy and speed is far beyond that of the algorithm in 2D image space.%基于图像/视频检测场景中的目标对象,总要涉及如何应对由成像透视造成的几何和尺度形变,以及运动形式变化等问题,使得在2D图像空间设计算法变得尤为复杂,尤其是针对目标遮挡和粘连问题,一直不能得到很好解决.为此,借助逆投影变换,提出一种在逆投影图中结合多部件混合模型的目标检测方法.首先在3D空间中构建与目标局部表面相贴合的逆投影面或阵列;然后,通过逆投影变换重构目标局部表面的特征数据,得到相应的逆投影图;接下来,提取部件样本在逆投影图中的HOG特征构建特征字典;并通过在字典上的稀疏逼近实现局部部件的检测.最后,利用部件检测结果,构建3D模型,并对质心聚类,完成最终的目标识别.实验表明,基于图像逆投影和3D部件混合模型检测图像/视频中的目标,不仅可以降低算法的复杂度,还可以从本质上有效解决目标遮挡和粘连,使检测精度和速度都远超于在2D图像空间设计的算法.
展开▼
