Improve the quality of ARC systems based on the metamorphic testing

In order to improve the quality of Activity Recognition Chain (ARC) systems, an effective testing approach is required to evaluate the functionalities of their major components, especially for two kernel components, the segmentation and recognition components. Both of them largely utilize the machine learning algorithms to achieve the detection and recognition of activity correctly. The probabilistic nature of the machine learning algorithms present severe challenge to the testing for the ARC systems. The article adopts a MT (metamorphic testing) methodology to address this topic. We analyze the MR (metamorphic relation) between the results of segmentation and recognition components. Based on this kind of MR, we propose a two-stage MT based approach for the testing of ARC systems, which exploits the association fact between segmentation and recognition, and then performs the MT testing on them respectively in sequence. This approach is able to deal with the specific issues in ARC systems testing, and provides a general method for systems testing with probabilistic nature.