针对视觉词袋模型与局部块模型特征相结合的算法在真实的复杂场景中识别率不高的问题, 本文提出一种基于局部块模型与特征预处理、特征泛化相结合的行为识别算法。本文算法在视觉词袋模型的基础上, 采取局部块模型与随机采样相结合的方法提取特征, 对特征做预处理, 减小了数据冗余, 消除了特征之间的相关性, 并且使处理后的特征更接近原始视频特征, 同时对编码后特征做泛化处理, 避免过拟合现象。本文在HMDB51标准视频库上进行实验, 结果表明本文算法较原算法识别率提高2.1%, 较其他同类算法也有一定的提升, 验证了该算法的有效性。本文算法对视频量大、背景复杂、真实场景的视频集具有较好的识别效果。

In order to solve the accuracy of the action recognition system in complex scenes, an improved method based on local part model, feature pre-processing and feature generalization is proposed. The algorithm is based on the bag of the visual words. Local part model and random sampling are used to extract features. The redundance of the features and the relationship between the features are reduced by the pre-processed procedure. The features processed are closer to the original features, meanwhile the features are generalized to prevent the over fitting. To demonstrate the performance of the proposed method, experiments are carried on HMDB51 datasets. The results show that the algorithm has higher efficiency than the original in complex environment. Compared with other methods, the proposed algorithm shows more accurate. It has a better recognition effect on video sets with large volume, complex background and real scene.