针对传统滑动窗目标检测方法需要在全图像范围内穷举搜索的缺点, 提出了一种基于视觉注意机制的粒子窗检测方法, 旨在保持较高检测精度的同时减少计算量。该方法将目标显著性作为先验知识引入搜索过程, 采用“图像签名”方法生成显著图, 然后通过阈值门限提取出包含有目标真实位置的局部区域。利用蒙特卡洛采样在显著目标对应的图像范围内均匀生成粒子窗, 并依据分类器的响应对粒子进行重采样, 以凸显真实目标区域、避免滑动窗方法对搜索步长的依赖。建立了Adaboost+类Harr特征(HLF)和支持向量机(SVM)+方向梯度直方图(HOG)的多级分类器结构, 前级分类器用于大范围目标的快速筛选, 后级分类器用于小范围目标的精确定位。将本文目标检测模型与传统滑动窗法和粒子窗法进行了比较, 结果表明本文方法的受试者工作特征曲线(ＲＯＣ)包含的面积更大, 耗时仅为滑动窗法的1/3到1/4, 粒子窗法的1/2, 在保持较高检测精度的条件下显著提升了检测速度, 实现了快速准确的目标检测。

As traditional sliding window detectors need to search the whole image by exhaustive method, a visual attention mechanism-aided target detection model by the particle window is proposed to reduce the calculational load while containing high detection accuracy. This model takes the target saliency as prior information of searching process, and then extracts the region of interest containing true target position by the “Image Signature” saliency map generator and entropy threshold. By uniformly drawing particle windows in an image range corresponding to the saliency targets with Monte Carlo sampling, the local region is treated as candidate detection points, thus resampling is carried out according to corresponding particle windows response. This strategy only focuses on the areas where the objects are potentially present and avoiding the tradeoff between accuracy and efficiency resulting from searching steps. A multi-stage classifier with Adaboost+HLF and SVM+ HOG is established, the former is applied to once-over and the latter is used to locate precisely. The target detection model proposed is compared with the traditional sliding window method and particle window method, and the results show that the Receiver Operating Characteristic(ROC) curve by proposed method contains the area to be larger than that of the other methods and the time consuming is only 1/3 to 1/4 that of the sliding window method and 1/2 that of the particle window method. It increases significantly detection speeds at maintaining high precision detection speed and achieves fast and accurate target detection.