Abstract: Most traditional methods for recognizing pornographic videos are extending the pornographic image recognition methods directly without considering  “behavior”, which is the pivotal information that pornographic videos contain. The optical flow context histogram can describe the continuous behavior of moving objects. For the above-mentioned reasons, this paper proposes a new feature used to describe the behavior—the Optical Flow Context Histogram（OFCH）, and uses Principal Component Analysis（PCA） for feature dimension reduction. The PCA-OFCH features are used to train sensitive action classifier. The use of Bayes skin color prediction model based on histogram techniques can determine if it contains sufficient color information in video, in order to reduce the false alarm rate. Experimental results show that the proposed features based on PCA-OFCH with skin color detection can effectively identify pornographic video and normal video, which provides new ideas for pornographic video recognition.

Key words: pornographic video, behavior detection, optical flow, Principal Component Analysis（PCA）, skin color detection

摘要： 传统色情视频识别方法大多是色情图像识别方法的直接扩展，没有考虑到“行为”这一包含在色情视频中的关键信息。光流上下文直方图能描述运动物体的连续动作，基于此，提出了一种新的用于描述行为的特征——光流上下文直方图（OFCH），并采用主成分分析（PCA）进行特征降维，得到的PCA-OFCH特征用于训练敏感行为识别器；同时采用基于直方图技术的贝叶斯肤色预测模型对视频中是否含有足够的肤色信息进行判断，以降低对正常行为的误报率。实验结果表明，提出的基于PCA-OFCH特征结合肤色检测能有效地对色情视频和正常视频进行鉴别，为色情视频识别提供了新的思路。

关键词: 色情视频, 行为检测, 光流, 主成分分析（PCA）, 肤色检测