Article Preview
Top1. Introduction
Service-oriented computing (SOC) is an emerging computing paradigm for integrating distributed services to meet dynamic business demands. Key elements of SOC include, among others, service-oriented architecture, service-oriented enterprise and web services (Kridel et al., 2017). Human flow counting has many applications in space management. For example, the flow count can be used to prevent over-crowdedness or to adjust HVAC settings accordingly in smart buildings. In order for SOC to work efficiently, each service end must be as autonomous as possible. Thus, a smart museum or concert hall must have some sort of automatic flow counting mechanism in order for its SOC based system to work. Mechanical units are commonly used to count human flow. However, they are inefficient and inconvenient. Image based solutions have been developed, but they require expensive camera devices and illumination and occlusion problems are unavoidable in image processing, let alone privacy issues.
The average human being contains about 60~70% water which disrupts radio wave propagation. Research on intelligent space management has shifted from image-based solutions to methods based on wireless technologies that are widely deployed in today’s smart societies. Wireless based solutions avoid the privacy invasion issue that often comes with image-based approaches.
Lin et al. (2011) exploited radio irregularity in the Internet of Things to count people automatically. The researchers used features extracted from received signal strength indicator (RSSI) to count the flow. RSSI, an aggregate power indicator resulting from the multipath propagation of indoor wireless communication (Yang et al., 2013), is widely available in many devices. RSSI is coarse-grained and more detailed information called channel state information (CSI) has been defined in the IEEE 802.11n standard (IEEE, 2009). Using hardware fast Fourier transform, off-the-shelf network interface cards such as Intel WifiLink 5300 (IWL) can output channel frequency response (CFR) as 30 CSI data per channel (Halperin et al., 2011). CFR is to RSSI what a rainbow is to a sunbeam (Yang et al., 2013).
CSI has a finer resolution than RSSI regarding communication channels, and exhibits a more stable temporal feature as well. Using empirical data, Wang et al. (2017) showed that CSI amplitude had a greater stability than RSSI for continuously received packets at a fixed location. The finer resolution comes with the price of higher dimensionality for CSI based features. Deep learning techniques were used to properly reduce the dimension and improve indoor localization accuracy (Wang et al., 2017).
Deep learning, a rejuvenated artificial neural network research subject, has caught the attention of many researchers in artificial intelligence. An essential part of deep learning is deep neural networks (DNNs) that can obtain useful features for object classification from raw image inputs (Hinton and Srivastava, 2006; Deng and Yu, 2014; Goodfellow et al., 2016; LeCun et al., 2015). It is thus expected that a well-designed DNN may be able to extract features from high dimensional CSI data to help the flow counting task.
In this study, the researcher exploited the opportunity of rich CSI data embedded in 802.11n networks to count human flow automatically. The counting problem was defined as a classification problem, i.e., using patterns of CSI fluctuation to predict the corresponding flow size. With 3 receiving antennas and 1 transmitting antenna, each network packet creates 90 CSI complex-valued data. Over a window of n continuously received packets, these 90n data were summarized into 630 inputs of magnitudes and 630 inputs of phases. Principal component analysis (PCA) is often used to extract or select features from high dimensional inputs before a classification algorithm is applied (Malhi and Gao, 2004). Instead of PCA, this study applied stacked denoising autoencoders (SdA) to extract hierarchical features from raw inputs (Vincent et al., 2008). On top of the SdA, a softmax regression was used to classify the last encoded features into different flow sizes.