Room occupancy evaluation through a single PIR-based motion detector and Machine Learning

This master thesis aims at creating a smart Pyroelectric Infrared based motion detector able to evaluate the room occupancy of an office. Since from the output signal of the motion detector it is not possible to directly evaluate how many people are present in the room, a counter of entrances and exits was used. In order to transform a motion detector into an entrance and exit detector, classical machine learning was exploited. Due to the limited time at disposal for the thesis, the PIR-based motion detector could only be mounted in one office and be analyzed during the spring season. For these reasons, it was not possible to record enough signal samples to generalize the results to another office or season. Thus, a procedure was established to automatically detect some entrances and exits in any possible room the detector will be mounted. Thanks to this rule, two one-class Support Vector Machines have been trained to detect entrances and exits starting from the ones following the automatic rule. Before training the two machines, the best possible PIR-based motion detector was searched for. Moreover, a long time was spent to define rules to extract the correct portion of the signal to be used in the training phase. Initially, the signal was divided into portions lasting a few seconds then, it has been divided according to its peaks. Another important step in the thesis has been the feature extraction and selection. Since the machines only understand numbers, the signal had to be described by numerical properties and, since the less number of features the easier the classification becomes, these properties had to be selected. Finally, the smart PIR-based motion detector has been tested on a real-life application and the results have been discussed in the conclusions section.