传统的基于特征提取的视频场景分类算法都是使用固定的量化方案,在量化的过程中对信息的利用程度不高.针对这一点提出了一种自适应的量化方案,对于光流的位置和方向分别进行非均匀的量化.这样能够最大限度地保留统计得到的底层特征分布信息,从而提高最终场景分类的整体算法性能.详细分析了自适应量化的原理,给出了自适应量化的步骤和算法流程,随后对比了使用改进的自适应量化方案和传统算法的实验结果.结果表明,改进后的算法一方面可以减少LDA模型的输入词典规模,提高运算效率,另一方面改进后的算法在检测的成功率上高于传统的算法,能够有效提升算法性能.%The traditional feature-based video scenes' classification algorithm usually use fixed quantifying methods, which do not fully use all the information in the process of quantifying. This paper proposed an adaptive quantifying method, which performs non-uniform quantization for both the location and direction of the optical flow. This method can maximize the retention of the characteristic information obtained from the featured-based statistical distribution lo improve the final performance of the algorithm. The theory of the adaptive quantifying algorithm is analyzed in detail, the steps are given. Then the results are compared with which just using traditional quantifying methods. The results show that the improved algorithm can reduce dictionary size of the LDA model and improve the operational efficiency. On the other hand, the improved algorithm has higher successful detection rate than the traditional algorithms, which effectively improves the performance of the algorithm.
展开▼
