Occupancy detection in the office by analyzing surveillance videos and its application to building energy conservation

Indoor occupancy measurement plays an indispensable role in occupant-based intelligent control of building systems for energy conservation. In this paper, a novel algorithm is proposed to detect occupancy by analyzing the office surveillance videos. The algorithm uses a cascade classifier to detect human head, consisting of pre-classifier, main classifier and clustering analyzer. The pre-classifier uses three frame difference algorithm to search motion windows and employs a HOG-SVM module to filter most non-head areas quickly. The main classifier employs a convolution neural network to classify head windows with high recall and precision. The clustering analyzer utilizes K-means clustering to fuse sequential frames and verify the head detection result to further improve the accuracy. The advantages of the three stages are enhanced by separately determined parameters and then united by the particular combination. The innovation yields an outstanding overall performance. The algorithm is tested on the dataset of 80-h surveillance videos in an office. The experimental results show that the accuracy (correctness for presence of head) of occupancy measurement reaches up to 95.3% and the computational cost for a measurement is just 721 ms. It is applicable to both off-line data mining of stored videos and on-line detection of occupancy by intelligent video surveillance.