This study aims to develop a decision procedure and associated small-sized, wearable wireless hardware module for successful fall detections of subjects for homecare. The detection procedure is developed based on a newly-proposed two-level criteria framework. The first-level criterion is forged based on two real-time signal vector magnitudes (SVMs) to determine whether falls can occur. The baseline SVM is further modified to accommodate varied gestures of the subjects prior to fall. It is designed that the modified SVM, named SVMB, with level exceeding 2.5 G indicates that the subject has fallen. The criterion at the second level is developed to prevent false detections, which considers especially the cases that some particular body movements other than falling would induce large SVMB values, such as running. Results based on a number of experimental trials validate the effectiveness of the proposed two-level criterion to detect falls while prevent false detections. A total of 100 experimental falls were tested, an average accuracy rate of 97 % in fall detection is achieved.