提出一种静态场景下的基于帧差光流的随机采样均值漂移聚类检测算法.该方法首先通过隔帧差分法提取运动目标区域, 并对运动区域进行光流计算, 采用自适应光流阈值分割法准确提取出运动目标; 然后运用连通区域标记算法对运动区域进行初步划分, 得到若干个连通域子集特征向量样本点, 通过提出的随机采样策略来确定对子集空间中样本点的抽样次数; 最后利用均值漂移算法对每个子集中的样本点进行若干次抽样计算并分析聚类收敛结果是否相同, 从而完成对连通域标记结果的检验.该策略通过减少对标记结果所有样本点的采样次数, 提高了目标的检测速度和精度,在不同红外测试场景中的实验结果表明: 与传统红外多目标检测算法相比, 该方法具有良好的局部抗遮挡性、准确性和实时性, 并且检测率能达到95.27%, 每帧处理时间达到39.12 ms, 满足实时处理需求.

A random sampling mean-shift clustering algorithm based on frame difference light flow was proposed. Firstly, the moving target region was extracted by frame difference method, and the moving region was calculated by optical flow, and the moving target was accurately extracted by adaptive optical flow threshold segmentation method. Then, the connected region labeling algorithm was used to preliminarily divide the moving region, and several connected domain subset eigenvector sample points were obtained. The sampling times of sample points in the subset space were determined by the random sampling strategy proposed. At last, mean shift algorithm was used to carry out several sampling calculations of sample points in each subset, and analyzed whether the clustering convergence results were the same. This strategy improves the detection speed and accuracy of the target by reducing the sampling times of all sample points of the marked results. Experimental results in different infrared test scenarios show that, compared with the traditional infrared multi-target detection algorithm, the method in this paper has good local anti-blocking, accuracy and real-time performance, and the detection rate can reach 95.27%, and the processing time per frame reaches 39.12 ms, which meets the real-time processing needs.