Abstract
Given the importance of recognizing indoor occupant's location and activity intensity to the control of intelligent homes, exploring approaches that can detect occupants' location and activity types while protecting occupants' privacy is helpful. In this study, occupants' zonal location and activity intensity recognition models were developed using passive infrared (PIR) sensors and machine learning algorithms. A PIR sensor array with 15 nodes was employed to monitor indoor occupant's and cat's behavior in a case residential building for 71 days. The output signals of PIR sensors varied with different locations and activity intensities. By analyzing the PIR data feature, models were established using six machine learning algorithms and two sets of data. After comparing model performance, the support vector machine (SVM) algorithm was selected to establish the final models. The model input was optimized by accumulating the PIR data. Taking PIR original counting values in 1-minute and 30-minute accumulated data as input features, the optimized SVM model can achieve 99.7% accuracy under the 10-fold cross-validation for the training data set, and 90.9% accuracy for the test data set. The cat's activity intensity is much weaker than that of occupant, yielding much smaller PIR output, which helped the SVM model to distinguish cat's activity from occupant's with > 90% accuracy. The model's recognition accuracy decreases with the decrease of sensor numbers and nine sensors were necessary. The findings obtained in this study support the promising future of applying PIR sensors in smart homes.
