Real-Time Continuous Activity Recognition With a Commercial mmWave Radar

mmWave-based activity recognition technology has attracted widespread attention as it provides the ability of device-free, ubiquitous and accurate sensing. Recognition of human activities intrinsically demands to be real-time and continuous, but the state of the arts is still far limited with the capacity in this regard. The main obstacle lies in activity sequence segmentation, i.e., locating the boundaries between consecutive activities in an activity sequence. This is a daunting task, due to the unclear activity boundaries and the variable activity duration. In this paper, we propose ZuMa, the first mmWave-based approach to real-time continuous activity recognition. When resorting to a machine learning model for activity recognition, our insight is that the recognition confidence of the recognition model is highly correlated to the accuracy of activity sequence segmentation, so that the former can be utilized as a feedback metric to finely adjust the segmentation boundaries. Based on this insight, ZuMa is a coarse-to-fine grained approach, which includes the fast coarse-grained activity chunk extraction and the find-grained explicit segmentation adjustment and recognition. We have implemented ZuMa with the commercial mmWave radar and evaluated its performance under various settings. The results demonstrate that ZuMa achieves an average recognition error of 12.67%, which is 65.08% and 71.87% lower than that of the two baseline methods. The average recognition delay of ZuMa is only 1.86 s.