基于目标检测及高密度轨迹的动作识别

Action Recognition Based on Object Detection and Dense Trajectories

为了实现准确的动作识别效果,我们通常需要提取能够充分代表运动特征的信息。近年来,基于高密度轨迹的动作识别方法因为能够提供丰富的时空信息而受到研究者们的广泛关注。但高密度轨迹类的动作识别算法通常都要面临背景冗余信息干扰的问题,为了解决这一问题,本文在高密度轨迹的动作识别方法基础上引入了目标检测算法,通过可变形块模型方法检测运动主体位置后计算其周围的高密度轨迹,有效地排除了背景冗余信息的干扰。而目标检测算法通常要面临丢帧问题,为了应对这一情况,本文采用了词袋模型和支持向量机进行动作特征表述和分类,词袋模型根据大量数据词频构建特征描述符的工作原理,使得目标检测偶有丢帧的情况并不影响动作识别的最终效果,结合高密度轨迹算法后有效地提高了传统高密度轨迹算法的效率,也获得了更为准确的识别效果。本文算法在KTH,UCF YouTube和UCF Sports数据集上较当前算法都取得了更高的动作识别准确率,尤其在复杂背景数据集UCF YouTube和UCF Sports上识别准确率分别可达89.2%和90.2%。

For recognizing human actions in video sequences, it is necessary to extract sufficient information that can represent motion features. In recent years, researchers pay more attention on dense trajectories because of its rich spatio- temporal information. However, dense trajectories based action recognition algorithms are all faced with redundant background problem. To ~lve this problem, we involve object detection in dense trajectories algorithm, detect motion object location through deformable part-based model and calculate dense trajectories around the motion object, which suppresses redundant background effectively. However, object detection algorithms are usually faced with missing frames problem. To solve this problem, human actions are classified by the bag-of-words model and SVM approach. Bag-of- words model constructs feature descriptors with word frequency, which makes few frames missing in object detection not influence action recognition result. Involving object detection improves dense trajectories approach efficiency, which also improves action recognition accuracy. Our algorithm achieves superior results on the KTH, UCF YouTube and UCF Sports datasets compared to the state-of-the-art methods, especially outstanding 89. 2 % and 90. 2% accuracy on complex background dataset UCF YouTube and UCF Sports respectively.