This section discusses about the related works for interpersonal distance estimation. As shown in previous works custom devices managed to achieve high accuracy. However, they have some major drawbacks including: a) requirement for specific on-body posi- tion of the device, b) need for additional hardware such as RFID Reader that must be
deployed in the experimental environment, which limits the mobility and the scale of the deployment, and c) users must carry non-pervasive devices, which affect their spon- taneous behaviour that is critical in this type of applications. To tackle these challenges the research target only commercial off-the-shelf mobile phones for the detection of so- cial interactions. For a more extensive analysis of the interpersonal distance estimation techniques the reader is refereed to Chapter 2.
3.2.1 Coarse-grained distance estimation
Literature initially focused on detecting social interactions through classifying nearby devices as interacting. Prior works like [46], [195] CenceMe [38], Serendipity [196] and SoundSense [45] performed discovery and the devices that were detected in vicin- ity were considered as in proximity. For the discovery process, researchers leveraged technologies such as Bluetooth, GSM and WiFi. Figure 3.2 depicts the users (A), (B), (C) and (D) in a social situation. Users (A), (B), (C) and (D) are detected as being in discovery range. It should be note that only users (A) and (B) are interacting in the particular context. Also, when the discovery range of a technology increases then the number of discovered devices may potentially increase in a multi-person situation when people are in a certain distance. Thus the number of false positives regarding the people socially interacting increases.
3.2.2 Fine-grained distance estimation
Estimating distance through only using the discovery range of a particular interface, introduces a large amount of error. To improve upon this approach, researchers fo- cused on reducing the discovery range of the interface. Literature avoided to modify the transmission strength of the signal as this would require firmware modification be- cause the mobile software development kits do not provide such functionality. Thus, to limit the range of the discovery, researchers induced various interpersonal distance estimation techniques. Among these techniques are the ToA, TDoA, AoA and RSSI. The first three techniques suffer from some significant drawbacks including a) the need for firmware modification, b) the requirement for external hardware, c) the need for
1.5m
3.0m
2.0m
3.5m
Figure 3.2: This figure shows the spatial arrangement of people including the con- cept of interaction zones [193]. In particular, the figure highlights the importance of understanding the interpersonal distance among people when socially interacting.
time synchronisation among the devices and d) the high development complexity. The last technique constitutes a more pervasive approach as RSSI is provided by the mobile software development kits, it does not fall in any of the above mentioned drawbacks and it has a very low development complexity.
Literature has focused on developing several path loss models based on Bluetooth RSSI such as Free Space PLM, Office PLM [143] and BlueEye [197]. Figure 3.3 shows the performance of the PLMs for interpersonal distance estimation based on the Bluetooth RSSI. This initial evaluation of the PLMs indicates the lack of ability to estimate interpersonal distance in real-world situations. This is because the PLMs are general analytical models and require specific parameters for a particular environment, but even in that case there is large deviation from the actual values. To further improve the interpersonal distance estimation, researchers developed a machine-learning technique called Comm2Sense [51] to perform proximity detection. They extracted the mean and the maximum of a 20-sample window of WiFi RSSI samples. The features were provided into a Na¨ıve Bayes classifier with a Kernel Density Estimator. As the approach operates in an ad-hoc mode, there was a need to switch between discovering and hot
spot mode, so the devices would be able to discover each other. Switching between the two mode is only possible through modifying the mobile phones’ firmware and it also induced a certain amount of delay in the discovering process. PhoneMonitor [49] is probabilistic proximity estimation technique that used Bluetooth RSSI to estimate how probable it is that the two people are in proximity to perform a social interaction. In MAUC [50] researchers developed a threshold-based technique to detect if two people are in proximity to interact considering also if they are standing or moving. The majority of these techniques are not able to cope with the RSSI fluctuation in real- world environments and do not consider human body absorption which is a critical factor for social situations. Also they perform the analysis off-line and not in real-time, introducing several privacy issues by transferring the data to a third party component and not performing the analysis on the users’ devices providing the with full control over their personal data.
In order to improve the accuracy of the distance estimation, researchers introduce a multi-modal approach Virtual Compass [53]. This method combines the WiFi and Bluetooth RSSI through developing regression models. The major shortcomings of this approach is the complexity of incorporating both modalities but also the high power consumption induced by simultaneous operation of both Bluetooth and WiFi. Both drawbacks indicate the lack of applicability of the approach in real-world social situations.
State-of-the-art techniques fall short mainly in terms of accuracy but also applicability in real-world environments. To address these challenges, in the next subsection an initial introduction of the DARSIS system is provided, an attempt to detect interpersonal distance through an accurate and reliable manner in real-world situations.
3.2.3 Proposed approach
In order to overcome the gaps identified in the state-of-the-art techniques DARSIS interpersonal distance estimation technique is designed, developed and evaluated. This technique estimates interpersonal distance in a fine-grained manner and detects the interaction zone in which users are and if they are in proximity or not. It performs the
Free Space PLM Office PLM [24] BlueEye [6] Baseline [10]% [11]%
Figure 3.3: This figure shows the performance of different propagation models for estimating distances 0.5, 1, ..., 4.0m based on Bluetooth RSSI.
detection opportunistically when a user is detected in vicinity to the smartphone and does not require any additional hardware or users’ involvement in the sensing or the inference process. This is a novel technique that requires only 6 Bluetooth RSSI samples to provide a fine-grained interpersonal distance estimation. Machine-learning models are trained for interaction zone and proximity considering the human body absorption, and show that they are able tolerate real-world fluctuation of the RSSI signal. The approach does not require any firmware modification and is able to operate real-time on an off-the-shelf smartphone, as it leverages the native ability of Bluetooth for ad-hoc discovery, allowing large-scale deployment through app stores. It is a privacy-preserving approach as the sensing and the inference are performed on-line on the smartphone and the data are not transmitted to any third party components.