Analysis And Research On Complex Action Recognition In Video

Complex action recognition is a hot research issue in the field of computer vision,which has a wide range of applications in intelligent video surveillance,video retrieval,human-computer interaction and other fields.The influence of complex backgrounds,diverse viewing angles,and changes within the class makes the recognition of complex actions challenging.At present,training a robust action recognition model requires a large amount of labeled data,and it is time-consuming and laborious to collect a large number of complex actions with tags.Considering that each complex action can be decomposed into a sequence of simple actions and these simple action can be easily obtained from existing datasets,we expect to help the learning of com-plex actions through the knowledge of simple actions.This article focuses on the analysis and research of complex motion recognition in video,the main tasks are as follows:1.Faced with the problem of having insufficient samples of tags,the knowledge of how to implement simple actions helps the learning of complex actions.To solve this problem,a simple to complex transfer learning model(SCA-TLM_H)is proposed.The SCA-TLM_H algorithm implements the transfer of knowledge from simple actions to complex actions using a matrix of prior knowledge.The relationship matrix encodes the transformation from simple actions to complex actions.The weight parameters of complex actions can be reconstructed by simple actions.Compared with the existing methods,the main advantage of SCA-TLM_His the use of simple actions to effectively implement knowledge transfer.To validate our approach,the proposed model was tested on two complex motion databases:the Olympic Sports dataset and the UCF50 dataset.The experimental results show the effectiveness of the proposed SCA-T LM_Htransfer.2.For the large number of types and quantities of actions in the SCA-TLM_Hmethod,designing the relation matrix based on prior knowledge is time-consuming and labor-intensive.It is very important how to adaptively study the relationship matrix.To solve this problem,a adaptive simple to complex action transfer learning model SCA-TLM_Ais proposed.In the proposed model,the relational matrix is obtained by minimizing the objective function.Through adaptive relation matrix optimizing the weight parameters,complex actions can be reconstructed by simple actions,and the target weight parameters are expressed as a combination of source weight parameters.Experiments are conducted in two complex action datasets,the results show that the adaptive learning relation matrix has more advantages.3.With simple actions,how to explore useful privileged information provides more in-formation during the training phase to help complex action recognition.To address this issue,a new learning framework was proposed,called Latent Task Learning with Privileged Infor-mation(LTL-PI).The probability matrix designed in this LTL-PI algorithm depends on label information and it is only available for the training data and is regarded as privileged informa-tion.The matrix encodes the probability of simple actions in complex action.By using the privilege information,the model is optimized and the best sparse weight parameter is finally obtained.The experimental results show that LTL-PI is an effective method to improve the recognition accuracy of complex actions,and the designed privileged information can provide useful help.