基于特定局部特征布局的视频运动目标挖掘

Video Moving Objects Mining Based on Distinctive Local Feature Configuration

针对局部特征描述子的高维特性以及视频中的复杂场景,提出一种利用频繁出现的特定局部特征空间布局对视频中运动目标进行挖掘的方法。在运动分割辅助下实现局部特征筛选,仅保留特定的运动相关特征,采用一种非参数维度约减方法建立精简描述子,并通过新的事务构建方式完成挖掘过程。标准数据集上的对比实验结果表明,该精简描述子在95%灵敏度情况下只有不到7%的假阳性率,整个挖掘方法相比同类方法具有更好的挖掘性能和可扩展性。

According to the requirements of coping with high dimensionality of local feature descriptors and clutter of video scenes, a new method of mining moving objects in videos using repetitive distinctive local feature configuration is proposed. Local features are filtered with motion segmentation aid, only the motion related descriptors are remained, and the condensed descriptors are constructed with a no parametric dimensionality reduction method. The final mining procedure is completed by a new transaction construction mechanism. Comparative test conducted on standard datasets demonstrates that the condensed descriptor gives 95% true positive at less than 7% false positive, and the whole mining method can exceed the mining performance and scalability of current state of the art techniques.