A similarity measure for analyzing human activities using human-object interaction context

Understanding the context of human-object interactions plays an important role in human activity recognition. Modeling the interaction context is a challenging problem due to the large number of possible objects in the scene and the large number of ways these objects may seem to relate to human activities taking place in the scene. In addition, providing labeling information of the object and human body parts is a very difficult and labor intense part of the training process. In this paper, we use a new class of kernels for image/video data as an extension of string kernels for 2 and 3 dimensional signals to model the human body parts and objects interaction context. In contrast to similar works, the proposed method does not require labeling of the human body parts and objects in the scene for the learning process, making it more practical when dealing with large datasets. Our experimental results show that the proposed kernel efficiently models the context of human-object interactions in image/video sequences and results in improved performance when compared to state-of-the-art methods.