Attribute-Based Transfer Learning for Behavior Recognition with Structural Information

Traffic behavior is closely related to the scene structure. Insufficient training samples will result in the problem of inappropriate scene modeling. An attribute-based transfer learning method for behavior recognition is put forward, which contains two main components: S-ATM Attribute Topic Model and S-CTM classifier model. S-ATM Attribute Topic Model is obtained by combined with the sources and sinks on the basis of the ATM Model, which realize modeling between behavioral properties and the underlying trajectory characteristics to learn the model parameters. Then the parameters of S-ATM Attribute Topic Model transfer to the S-CTM classifier model as attribute prior knowledge to improve the performance of target behavior classifier model. Experiments show that the proposed approach can transfer attribute priors from source categories to target categories effectively and significantly improve the recognition accuracy rate of the target behavior classifier model in zero-shot learning task.