Research On Action Recognition Based On Spatio-temporal Trajectory Matching

Action recognition has a wide range of potential applications in many areas,such as intelligent video surveillance system,video retrieval,human-computer interaction,it has become a research hotspot in computer vision.Although there have been numerous works on this respect,it is still one of the challenging problems in computer vision.First,there is a huge intra-class change caused by the difference in motion speed,viewpoint change and background noise.Second,The definition of a class is based on the concept and semantics of the action,the deviation between the low-level video feature and the high semantics makes the classification very difficult.There has been a lot of excellent literature in action recognition,among them,dense trajectories as a highly robust low-level feature and Fisher Vector are widely used as action recognition as a common solution,and shows the state-of-the-art performance on several popular benchmarks.In this solution,Gaussian Mixture Model(GMM)based codebook is built to represent the distribution of local features in training videos,which is then used to encode the local features of a given video.In this approach,all trajectories are encoded disorderly,while the relationship between different trajectories is ignored.In order to reduce the loss of spatio-temporal information.In this paper,two different trajectories matching strategies are proposed to capture the lost spatio-temporal information,termed as KNN-based feature matching strategy and stacked feature matching strategy.Firstly,we define a trajectory distance to measure the relationship between the two trajectories.Then,perform the feature matching strategies proposed in this paper to match the trajectories.For the pairwise trajectories matching,we use the average pooling strategy to carry out the feature fusion to get the space-time pairwise trajectories(SPT)proposed in this paper.A GMM is then trained and Fisher Vector is employed to quantify the pairwise trajectories.In this way,the local spatial and temporal structure information around each trajectory is explored for the final representation,which improves the discriminative ability of the features.The experimental results on four benchmark datasets(Olympic Sports,HMDB51,UCF50,UCF101)show that our pairwise trajectory representation outperforms the state-of-the-art approaches.