多尺度局部区域响应累积的非滑窗快速目标检测算法

针对滑动窗口全局搜索检测目标搜索时间长的问题，提出一种多尺度局部区域响应累积的非滑窗快速目标检测算法。首先，提取检测目标多尺度可重叠局部区域作为训练样本，通过学习得到多尺度且具有判别能力的部件集，部件集中每个局部区域与检测目标有明确位置对应关系；然后，根据各投影检测器响应判断目标是否在某一区域出现，并利用多尺度目标局部区域的检测结果和位置约束进行投票，完成目标粗定位；其次，利用HOG特征提取和SVM相结合完成检测目标验证实现准确检测。该方法将多尺度部件模型、统计累积投票思想及分类器判决相结合，实现快速目标检测，大大减少滑动窗口逐像素搜索背景时所消耗时间，提高检测效率。 

Fast object detection algorithm with non-sliding window based on accumulation of multi-scale local response

Exhaustive search method of sliding window which consumes much time in searching the location is used to detect objects. In order to solve this problem, we propose a fast object detection algorithm with non-sliding window based on accumulation of multi-scale local response. Firstly, the multi-scale and foldable local areas are extracted as the training sample, then learning them to obtain a part sets with multi-scale and discriminative ability, in which every local area and object have a definely position corresponding relationship; Secondly, the appearance of a particular area is based on every projection detector’s response and the object’s positions are determined by using voting shceme with multi-scale object detection results of the local area and the position constraint; Finally, we test the object by combining the HOG feature extraction with the classifier of SVM to realize accurate location. Experimental results show that the proposed method which combining multi-scale part model and statistics of cumulative voting and the classifier of SVM improves the detection efficiency via saving the consumed time of the sliding window pixel-by-pixel searches background.