Action Detection Based on Latent Key Frame

Human action detection in videos is a challenging problem in the field of Computer Vision and it has become an active researching field in recent years. For most published methods, which analyses entire video and assign a single action label; by contrast, in our research, it has been proved that most of actions could be detected within only a few frames. Based on this hypothesis, a temporal structure based model named Latent Key Frames Model (LKFM) is proposed, in which the action was represented as a sequence of Key Frames. LKFM is able to find the optimal Key Frames sequences with the help of latent support vector machine (Latent SVM); and for each Key Frame in the Key Frames sequence, a 2d model is built with the help of Deformable Part-based Model (DPM). The proposed method has been evaluated on Weizmann dataset and UCF sports dataset, and the experimental results demonstrate that this model is able to achieve competitive performance.