Autonomous Human Flow Counting Service With Deep Neural Network

Abstract

Human flow counting has many applications in space management. This study applied channel state information (CSI) available in IEEE 802.11n networks to characterize the flow count. Raw features including the mean, standard deviation and five-number summary were extracted from CSI magnitudes in a time window. Due to the large number of raw features, stacked denoising autoencoders were used to extract higher level features and a final layer of softmax regression was used to classify the flow count. The resulting neural network beat many popular classification algorithms in predicting the correct flow size. In addition to CSI magnitudes, this study also explored the feasibility of using CSI phase-based features. It is found that the magnitude neural network provided a better prediction result than the phase neural network, and combining both networks yielded an even better solution to the flow counting problem.