一种融合预测过采样的运动目标检测算法

为了改善运动目标检测的精度,提出了一种融合了预测过采样的运动目标检测新方法.首先,基于二维傅里叶变换预测当前帧的目标形状并计算形状相似度;然后,从历史检测结果中选择一定数量的参考帧,使用光流法跟踪目标像素点在参考帧与当前帧之间的运动轨迹,并以像素点轨迹为参考在采样区间执行稠密过采样;最后,基于过采样样本构造前景模型,并在图分割框架内联合使用前景背景模型实现目标检测.在公共数据与自采数据集上对所提方法进行了实验验证,结果表明,相对于经典的运动目标检测算法,所提方法能够有效提高检测精度.

Prediction based over-sampling for moving object detection

Prediction based over-sampling is proposed for moving object detection. First, 2D Fourier trans-form is used to select some historical foreground frames as references of over-sampling, where the selection rule is shape similarity. Then foreground pixels are tracked from reference frames to the current frame with optical flow method. Next, dense over-sampling is conducted to generate new samples which are then used for improved foreground modeling. Finally, the foreground model is used in conjunction with another back-ground model for graph cut based classification. Comparison between the presented method and state-of-the-arts shows that the generated samples are representative and the proposed algorithm is effective.