Abstract: In order to solve the issues in the two-person interaction recognition algorithm, such as the high dimension of the feature and weak representation ability, a novel algorithm based on the active curve model is proposed in this paper. Due to the advantage of the active curve to the sparse representation ability of the targets, the proposed algorithm obtains the deformable templates of the active curve by utilizing Gabor filter and Sum-Max Maps, then gives the sparse representation of two-person interaction. In addition, the proposed algorithm uses HOG features to describe all frames in the video sequences, then innovatively uses the distance of the extreme value method to get key frames of video sequences. Finally, the performance of the proposed algorithm is tested on the UT-Interaction dataset. The experimental results show that the active curve model extracted in the key frames is simple and has better representation ability, which obtains better interaction recognition rate. So the sparse representation of the algorithm in the field of interaction recognition has good research prospect.

Key words: Interaction recognition, sparse representation, active curve, key frame

摘要： 针对双人交互行为识别算法中特征维数过高且表述能力不强的问题，提出一种基于活动曲线模型的双人交互行为识别算法。该算法利用活动曲线对前景目标稀疏表示能力强的优点，采用Gabor滤波和Sum-Max Maps的方法得到活动曲线的可变模板，进而对双人交互行为视频中的关键帧进行稀疏表示。该算法中提出利用HOG特征来描述视频序列中每帧图像，然后利用距离极值得到视频的关键帧的新方法。在UT-Interaction数据库上的测试表明，在运动视频的关键帧中提取的活动曲线模型特征简单，具有较好的行为表述能力，得到了理想的双人交互行为的识别准确率，充分验证了稀疏表示算法在双人交互行为识别领域有较好的研究前景。

关键词: 交互行为识别, 稀疏表示, 活动曲线, 关键帧