Abstract: To improve the accuracy of group activity recognition in video, a group activity recognition algorithm was proposed based on tensor feature and twin support vector machine. Firstly, the activity of group in each frame was described by combining the posture structure information in the joint skeleton of the group members and the social network information of the group. The tensor form was used to represent the features of group activity. Then, the tensor kernel was decomposed by using multi-channel nonlinear feature mapping and the model parameters of the tensor kernel twin support vector machine were optimized by using the particle swarm optimization method. Finally, the group activity recognition in video was realized by combining tensor features and twin support vector machine. Experiments performed on the CAD2 dataset and the self-built dataset show that the tensor feature can effectively represent the group activity. Compared with the existing approach, the proposed algorithm can effectively improve the accuracy of the group activity recognition.