复杂背景下目标检测的级联分类器算法研究

目标检测与跟踪一直是图像处理与计算机视觉领域的热门研究方向之一,其对军事上的成像制导、跟踪军事目标等以及民事方面的安防监控、智能人机交互等方面均有着重要的研究价值.将特征匹配问题看成是一种更普遍的二分类问题,将这种难解的高维计算变成二分类问题,使计算复杂度大大减小,这类方法以大数定律和贝叶斯法则为理论依据,本文提出一种非树形结构的分类器,并从理论上推导出其实现公式,将1bitBP特征应用到分类器中,同时采用计算量由小到大的三个分类器进行级联从而实现鲁棒精确的目标检测.从实验结果来看,本文算法能够对目标的尺度变化、旋转、部分遮挡、形变、模糊、背景变化等复杂情况有较好鲁棒性,并且检测精度相对较高,而本文算法的计算复杂度低、计算量小,有较高的应用价值.

Study on the cascade classifier in target detection under complex background

Method of target detection and tracking is one of the hot topics in image processing and computer vision field, which is significant not only in military such as imaging guidance and military target tracking, but also for civil use such as security and monitoring and the intelligent man-machine interaction. Treating the feature matching problem as a more general equinoctial classification question, can turn the intractable high-dimensional problem to a classification problem and deplete computer complexity. This method is based on the law of large numbers and Bayes rule. In this paper we propose a non-hierarchy structure classifier, for which the equation for calculation is theoretically derived, and apply 1bitBP feature to the classifier; and for further reducing the amount of calculation, we use integral image and square integral image to variance classifier as preprocessor, and then use non-hierarchy classifier to handle the patches which meet the variance demand and use the nearest neighbor to further improve the accuracy, and finally realize target detection and tracking based on cascade classifier. Our experimental results show that the method proposed is far superior in calculation amount and processing precision, and is robust to scale changing and rotation, so the method proposed in this paper is of high practical value.
Keywords：target trackingtarget detectionran treeclassification