Projects:2017s2-201 Detection and Classification in an Indoor Environment Using WiFi
- 1 Introduction
- 2 Related Work
- 3 Background
- 4 The Experiment
- 5 Conclusion
- 6 Future Work
- 7 Project Team
- 8 Supervisors
Indoor detection and localization is an area that has recently received attention. Whilst several device tracking solutions have been achieved, there is little work that has been conducted in the area of device-free detection. Device-free detection includes intruder detection, path tracking, and object location amongst other things. There have been a variety of approaches to solve this and other similar problems, including Received Strength Signal Information (RSSI), Bluetooth, passive radar and GPS. Whilst technologies such as GPS and Bluetooth are able to accurately determine the location of a device in relation to its surroundings, there is a huge gap in the ability to detect changes in an indoor environment that are unrelated to a device. This project aims to produce a system that is able to achieve accurate device free detection and localisation using Wi-Fi.
Wi-Fi is a readily available, ubiquitous technology that many people use daily. It allows people to be connected at all times, communicating and learning at all times, from all corners of the globe. Wi-Fi packets each carry a multitude of information, one aspect of which is channel estimation, found in the physical (PHY) layer. It describes the propagation of a signal between the transmitter and receiver, or more colloquially the channel between the two.
Channel estimation is represented by a series of subcarriers produced through Orthogonal Frequency Division Multiplexing (OFDM), and can be used to describe the characteristics of the environment surrounding the signal path. The number of subcarriers that are present in a packet is dependent on the operational frequency. The channel estimation, or Channel State Information (CSI) is represented as a complex number, the values of which cluster due to different signal paths. CSI is a source of fine grained channel information, in that the channel is estimated using multiple values rather than just a single value.
This project aims to develop an accurate indoor object detection system using Wi-Fi, through the analysis of CSI data. It consists of two major focus streams: hardware analysis and software analysis.
- Development of fingerprints through use of multiclass SVM
- Investigation into computational efficiency and prediction accuracy
- 2.4GHz against 5GHz frequency bands
- single transmit stream against two transmit streams
- single packets against multiple packets as a data input
- Development of 3 systems for CSI collection
- Laptop-laptop pair using the Intel 5300 CSI Tool
- Laptop router pair using the Intel 5300 CSI Tool
- Router router pair using the Atheros CSI Tool
- Investigation into the performance of the various hardware configurations under different testing conditions
- Testing conditions include:
- Unique object identification
- Unique location identification
- Simultaneous object and location identification
- Testing conditions include:
A number of different groups have previously investigated using Channel State Information (CSI) to perform intruder detection, device localization and even activity identification. The CSI values are used by these groups to detect changes in the environment, with changes being identified through use of a trained machine learning algorithm, often a Support Vector Machine (SVM). The training process allows for unique characteristics to be identified and associated with a particular configuration.
Whilst the complexity of these previous studies varies, the results from each suggest that CSI is able to accurately detect changes, implying that the proposed project is feasible.
Device Free Detection
Device-free detection is the process of identifying a sensor free target. This target can be passive (stationary) or static (moving), and any size. This method of detection poses many challenges, as communication cannot occur between the target and the detecting unit, causing typical detection methods such as Bluetooth or GPS to be superfluous. Current systems that can achieve device free detection rely upon electronic systems such as cameras or inbuilt sensors to a room, both of which are costly in terms of dollar value and hardware maintenance requirements. There is hence a market for a cost efficient system to achieve this, and this project looks at developing one such system.
IEEE 802.11 standards provide the specification for Wireless Local Area Network. One or more of the 802.11g/n/ac protocols are used on all current commercially available routers. This project concentrates on the 802.11n protocol.
The 802.11n protocol operates on two frequency bands, the 2.4GHz and 5GHz bands. In the 2.4GHz frequency band, there are either 14 channels, 13 of which are usable in Australia, each channel with bandwidth of 20MHz, or 9 channels, each with channel bandwidth of 40MHz. In the 5GHz frequency band, there are 42 available channels, each channel with bandwidth of 20MHz/40MHz. Most channels overlap each other, resulting in interference, however there are some channel combinations that do not interfere. Each channel has multiple subcarriers that carry the signal simultaneously. The 802.11n protocol employs Orthogonal Frequency Division Multiplexing (OFDM) as a digital modulation technique. The data is modulated on multiple subcarriers in different frequencies, with all subcarriers orthogonal within the same channel in order to avoid interference between each of the subcarrier frequencies.
Channel estimation is a way of describing the channel that exists between a transmitter and a receiver. The channel that is described will vary as the surrounding surfaces are changed, whether new surfaces are added (eg a table is introduced into a room) or old surfaces are removed. This variation is caused by the change in reflective surfaces for signals to reflect off of, altering the transmission path(s) taken by the signal between the transmitter and receiver.
There are two main methods used for channel estimation: the Received Signal Strength Indicator (RSSI) and Channel State Information (CSI). RSSI provides a single real value to estimate the channel, where as CSI provides a series of complex values, one value for each of the subcarriers from the modulation. As such, RSSI is a coarse grained estimation method and CSI is a fine grained estimation method.
This project utilises the fine grained CSI.
There are several groups that have devised tools to allow for CSI extraction. Two of these tools are the Intel 5300 CSI Tool, developed by D. Halperi and the Atheros CSI Tool, developed by Y. Xie. These two tools extract the (usually unavailable) CSI data directly from received packets. The Intel tool is able to extract the CSI from 30 of the subcarriers (effectively one in two on the 2.4GHz band) and the Atheros tool is able to extract the CSI from all 56 (of the 2.4GHz band) of the subcarriers. Both tools operate on the 802.11n standard, and require transmission at a high threshold rate.
Fingerprints are data collected from experimental measurement in a controlled environment to build a radio map. Passive detection uses fixed AP, test object is moving in the testing environment to create fingerprint. When the test object is introduced in an unknown location in the testing environment, by comparing the similarity of the fingerprint, the estimated location could be identified. In this project, CSI data are collected in 15 different locations arranging in a grid. Each object is placed in all 15 locations to collect CSI. After CSI collection and extraction, machine learning algorithm is used to detect and classify the object.
Machine learning is an algorithm that develops hypothesis by utilising a known set of examples called training set. Through search of possible hypothesis, an input to the machine learning algorithm will result in the probability of target output closest to the hypothesis. In the project, Support Vector Machine (SVM) is used as the machine learning algorithm. SVM select a smaller subset of instances to form decision boundary. A linear discriminant function is used to set the separation margin between two classes as wide as possible, usually by transforming the data into higher dimensional space.
To further add to the challenge, the environment considered by this project is an indoor one. Indoor environments present additional difficulties, including multipath, scattering, fading, and delay distortion. Each of these has a unique impact on each signal that transmits through the environment, making the transmission inconsistent and, at times, unreliable.
Two laptops equipped with Intel5300 network interface card are used as transmitter and receiver pair. A chair, a desk and a human were placed in turns in each of the 15 locations (layout as seen in below hardware section) for Wi-Fi packet collection. Packet collections are done in two different transmission settings:
- transmitter set as 2.4GHz with 20MHz channel bandwidth
- transmitter set as 5GHz with 20MHz channel bandwidth
The flow for data collection and analysis is shown above. CSI values are extracted after processing the packets collected. The CSI values formed a set of set. These set of data are split into training dataset and testing dataset. Training dataset is then used to build machine learning models, whereas the testing of the models is done by utilising the testing dataset.
The hardware aspect of this project had two phases: system development and data collection/processing.
Three systems were investigated for development, two based on the Intel 5300 Tool and one on the Atheros CSI Tool. The systems are briefly summarised as follows:
- Laptop pair operating the Intel 5300 Tool
- Laptop router pair operating the Intel 5300 Tool
- Router pair operating the Atheros CSI Tool
Both the Intel and Atheros tools had several system requirements:
- Operating on an unencrypted network
- Network operating on the n Wi-Fi standard
- Transmission rates being of High Throughput
The first of the Intel systems was provided to the project team by the Defence Science and Technology Group. It consisted of two Lenovo X200 laptops operating Intel 5300 Network Interface Controller (NIC) cards.
The second of the Intel systems was developed throughout the project. Its final form consisted of a Lenovo X301 laptop paired with a router (specifically, a Wi-Fi Pineapple). The laptop had a Intel 5300 NIC card installed, and a network was established by the router and connected to by the laptop, establishing a connection. Issues that were experienced with this set up included driver installation and kernel usage, as both had specific requirements. The driver installation order provided particular difficulty as rebooting the computer would change the order in which they were installed without any warning.
The third system was unable to be successfully built due to complications. The Atheros system required the installation of OpenWRT onto a router, and also required the router to operate on an Ath9k NIC that supported beamforming. These specifications were challenging to meet with currently available commercial routers, as most current routers that have the correct chipset are later versions, meaning that they do not support updating firmware. Several routers were purchased but upon arrival it was discovered that they were unsuited to the task due to later versions being delivered, and earlier versions being unavailable for purchase.
Data Collection and Processing
Data was collected on the two functional systems for three unique objects and 15 locations. The image below shows the layout of the room in which data was collected, with X's showing the location of each of the collection points.
The data that was collected at each of these points was used to train a variety of models for different testing purposes, detailed in the following.
- Testing if an object is present or absent
- Testing what object has been introduced
- Testing where an object has been introduced on coarse and fine grids
- Testing what object has been introduced and where it has been introduced on both coarse and fine grids
Tables 1 and 2 provide a representation of how the two systems perform in a basic case. The trend of System 2 outperforming System 1 was consistent throughout all testing.
|System 1||Predicted State|
|System 2||Predicted State|
From the investigation that was conducted, it was discovered that the second of the two Intel systems (the laptop router pair) was superior at detecting and classifying changes to a room. This was shown by a significant increase in accuracy of prediction when comparing the two systems directly.
There are several paths that future work on this project could follow, listed below.
- Complete the development of System 3. This would allow for direct comparison between the two tools (Intel and Atheros) which, from our knowledge, has not yet been done.
- Exploration of data processing methods to increase detection accuracy
- Expansion of system to a broader environment. This could take several forms, including
- Different rooms
- Ability to identify a larger range of objects
- Potential to provide a size approximation of the object
- Ability to identify the location more accurately
- A system not limited to specific locations, rather one that functions more like a grid in terms of location prediction
- Tracking of a moving object
The exploration of these paths would allow for a more accurate and robust system to be developed for use in many applications.
Tin Yan Wong
Cheng Chew Lim
Hong Gunn Chew
John Kitchen (DST Group)