Nowadays, Coronavirus disease (COVID-19) has become a global pandemic because of its fast spread in various countries. To build an anti-epidemic barrier, self-isolation is required for people who have been to any at-risk places or have been in close contact with infected people. However, existing camera or wearable device-based monitoring systems may present privacy leakage risks or cause user inconvenience in some cases. In this paper, we propose a Wi-Fi-based device-free self-quarantine monitoring system. Specifically, we exploit channel state information (CSI) derived from Wi-Fi signals as human activity features. We collect CSI data in a simulated self-quarantine scenario and present BranchyGhostNet, a lightweight convolution neural network (CNN) with an early exit prediction branch, for the efficient joint task of room occupancy detection (ROD) and human activity recognition (HAR). The early exiting branch is used for ROD, and the final one is used for HAR. Our experimental results indicate that the proposed model can achieve an average accuracy of 98.19% for classifying five different human activities. They also confirm that after leveraging the early exit prediction mechanism, the inference latency for ROD can be significantly reduced by 54.04% when compared with the final exiting branch while guaranteeing the accuracy of ROD.