Fine-grained Indoor Localisation using Wireless Sensor Networks. Katelijne Vandenbussche

Transcription

1 Fine-grained Indoor Localisation using Wireless Sensor Networks Katelijne Vandenbussche

2

3 Fine-grained Indoor Localisation using Wireless Sensor Networks Master s Thesis in Computer Science Parallel and Distributed Systems group Faculty of Electrical Engineering, Mathematics, and Computer Science Delft University of Technology Katelijne Vandenbussche 17th August 2005

4 Author Katelijne Vandenbussche Title Fine-grained Indoor Localisation using Wireless Sensor Networks MSc presentation 25th August 2005 Graduation Committee Prof. dr. ir. H.J. Sips (chair) Dr. K.G. Langendoen Dr. A. Baggio Dr. D.V. Keyson Delft University of Technology, Faculty of Electrical Engineering, Mathematics and Computer Science Delft University of Technology, Faculty of Electrical Engineering, Mathematics and Computer Science Delft University of Technology, Faculty of Electrical Engineering, Mathematics and Computer Science Delft University of Technology, Faculty of Industrial Design Engineering External supervisor ir. M.H. Vastenburg Delft University of Technology, Faculty of Industrial Design Engineering

5 Abstract This thesis describes the research conducted on algorithms for wireless sensor networks that are applicable for fine-grained localisation in a small indoor environment. The goal of our project is to provide location information about mobile sensor nodes that is accurate enough to be used for activity recognition of the persons carrying the nodes. The scalability of existing localisation algorithms, designed for outdoor and large indoor environments, is investigated. Our preliminary experiments show that algorithms based on radio signal strength should provide satisfactory results for small-scale indoor localisation. Two localisation systems based on radio signal strength were therefore designed and implemented: a fingerprinting system and a proximity-based system. The experiments that were conducted to test the systems show that the localisation errors achieved by the fingerprinting system are variable, and therefore that the system is unreliable, whereas the proximity system produces accurate, relative location information. Consequently, only the proximity solution is applicable for the activity recognition problem, although minor improvements to the system are still necessary.

6 iv

7 Preface The report you are about to read is the result of my Master of Science thesis, which I carried out in the Parallel and Distributed Systems group of the Computer Science department at Delft University of Technology, in collaboration with the Faculty of Industrial Design Engineering (IDE). It is the result of a challenging project, which took me about eight months to complete. The reason for choosing the subject of wireless sensor networks for my thesis is that it is a rather emerging and challenging research topic. There is still much to be found out, which makes the subject very interesting, but on the other hand also frustrating sometimes. Furthermore, the idea of having to cooperate with people of IDE motivated me for doing the project. These people are known to be much more open-minded and creative than the average computer scientist. The latter surely was an attractive idea after all these years of pure technical thinking. Before I continue with my report, I would like to thank some people for their support during my work. In the first place, my thanks go to Siebe, who has been supporting, patient and inspiring all the time. Of course I m also grateful to my parents for their interest and financial support during the past five years. I also want to thank Pixel, my cat, because I could always cry out on her and tell her my problems and frustrations. I would like to thank my daily supervisors, Martijn, Aline and Koen, for their support and constructive comments during our regular meetings. My last thanks go to my friends and my colleagues at the 9th floor, who were always willing to help me, to inspire me and last but not least to entertain me with their exciting and unbelievable stories. Katelijne Vandenbussche Delft, The Netherlands 17th August 2005 v

8 vi

9 Contents Preface v 1 Introduction 1 2 Problem Settings and Requirements Project Description Project Objective System Requirements Related Work Localisation in WSNs Definitions Existing Techniques and Algorithms Localisation using GPS Localisation using Infrared Localisation using Sound Radio-based localisation Summary A Preliminary Study of Applicable Techniques Available Hardware and Software Chosen Localisation Technique Indoor Characteristics of Radio Signal Strength The Reproducibility of RSSI measurements Relation between RSSI and Distance Influences of a Dynamic Environment Conclusions Two Different Localisation Systems The Fingerprinting System Motivation System Setup and Design Implementation Details TNOde Software vii

10 5.3.2 Computer Software Experimental Results Self-localisation System Performance System Review The Proximity System Introduction Motivation System Setup A Preliminary Experiment System Design Implementation Details TNOde Software Computer Software Experimental Results Tuning the Proximity System The System s Lifetime The System s Performance System Review Conclusions and Future Work Conclusions Future Work A A Detailed Description of the Proximity Code 51 viii

11 Chapter 1 Introduction The Master s project described in this report is part of the research on Wireless Sensor Networks (WSNs) that is currently conducted at the Computer Science department at Delft University of Technology. A wireless sensor network is a network that consists of numerous small devices, that are in fact tiny computers. These so-called nodes are composed of a power supply, a processor, different kinds of memory and a radio transceiver for communication. Wireless sensor networks are generally used to observe or sense the environment in a non-intrusive way. In order to perform this task, nodes are often extended with sensors, like infrared or temperature sensors, hence the names sensor nodes and sensor networks. The domain of wireless sensor networks is still very young. During the last few years, new developments in the area of communication, computing and sensing have enabled and stimulated the miniaturisation and optimisation of computer hardware. These evolutions have led to the emergence of wireless sensor networks. Despite the increasing capabilities of hardware in general, sensor nodes are still very restricted devices. They have a limited amount of processing power, memory capacity and most importantly energy. Therefore, most of the existing operating systems, programs and even algorithms are not suitable for WSNs. Their memory and processing requirements are too high given the constraints of sensor nodes, and power consumption would be exhaustive. This makes wireless sensor networks a challenging research topic. The main research target is the overall reduction of power consumption. Therefore, research is done on the implementation of small and capable operating systems and on the construction of efficient algorithms, in terms of energy, processing power and memory footprint. Despite current restrictions, several applications for wireless sensor networks have already been designed. Wireless sensor networks are currently found in very different domains [7, 16]. In the medical world, they are used to improve the process of treating and following-up patients. The military world uses wireless sensor networks in hostile environments to spy on the enemy in a safe, non-intrusive way. 1

12 Wireless sensor networks are found in civil applications too, for monitoring and protecting public property. In control and automation applications, they aim at monitoring and improving industrial processes and in safety and security applications, they try to protect people and private property from malicious attacks. WSNs are also used for environmental monitoring, which is the long-term observation of natural phenomena, fauna and flora, in a non-intrusive way. The observation data is studied and measures can be taken to adapt the environment if necessary. Finally they are found in home applications, to ease people s life by making the home environment smarter in any possible way. When talking about smart home environments, one could in the first place think of obvious functional applications, like the lights being turned on automatically when it gets dark or the air conditioning being turned on when it is getting too warm. At the faculty of Industrial Design Engineering (IDE), researchers want to exploit the possibilities of WSNs even further. Their goal is to create complete atmospheres, in which the settings for among others light and music are defined. These atmospheres can subsequently be linked to activities. One could for example think of a warm atmosphere, with quiet music and dim lights, for intimate or familial activities. A test setting for this atmosphere creation process, a laboratory living room where people can select atmospheres using an interactive interface, already exists. The researchers at IDE would now like to automate the process: atmospheres have to change automatically based on knowledge of people s locations and activities. For this purpose, a WSN can be used. Sensor nodes are carried around by people and a fixed network infrastructure is used to track the mobile nodes. This process results in detailed location and mobility information about the persons in the room, which can subsequently be used for activity recognition purposes. The project illustrates the need for fine-grained indoor localisation using wireless sensor networks. According to the literature, this is a difficult problem [3, 12, 14, 19]. Efforts are made to achieve accurate localisation in WSNs using a wide variety of technologies, such as infrared, ultrasound, GPS or radio signal strength indicator (RSSI), but most of the current algorithms offer coarse-grained localisation only. However, a positive trend can be seen in the results over the last few years, due to improved technologies. In this thesis, we will try to improve two of the existing algorithms even more. The localisation systems presented in this report are based on the radio signal strength indicator. The first solution is a fingerprinting algorithm. It relies on a map of measured signal strengths at predefined locations in the room, called fingerprints. Run-time RSSI values taken at the node s current location in the room are compared to the fingerprints to find the most probable location. The next technique is a proximity solution. Sensor nodes that are used as reference points are spread around the room. RSSI is measured between the reference points and the node to be located and it is used to determine the closest reference points at any time. By analysing the variation in measured signal strengths over time, we can also find 2

13 out whether a person is moving or not. Based on this information, a person can be tracked and broadly located. Finally, the location information will be used by researchers at IDE to recognise the people s activities. If, for example, all people are located close to the dining table and they are not moving, it is likely that they are eating. The atmosphere is then automatically adapted accordingly. The remainder of this report consists of six chapters. In Chapter 2, general information is given about this Master s project. The context of the project is described, followed by a more detailed description of the problem, the objectives and the requirements. In Chapter 3, an overview of the research that has been done up to now with respect to localisation using wireless sensor networks is presented. This research forms the basis for the solutions for this thesis project. Chapter 4 introduces the implemented solutions. The choices are motivated and preliminary experimental results are presented. In Chapters 5 and 6, the fingerprinting solution and the proximity system are presented, respectively. The details about the setup, design and implementation of the system and experimental results reviewing them are given. In Chapter 7, conclusions about the project are drawn and recommendations for future research are given. 3

14 4

15 Chapter 2 Problem Settings and Requirements This chapter presents general information about the localisation project. In Section 2.1, the context of the project is given and the concrete problem that is the reason for doing this research is presented. The objective of this Master s project is stated in Section 2.2. Section 2.3 lists the specific requirements for the localisation system to be designed. 2.1 Project Description The StudioHome lab [13] (see Figure 2.1) is an artificial living room that is set up at the Faculty of Industrial Design Engineering (IDE) at the Delft University of Technology. The room offers a way of changing environmental factors, such as light intensity and colour, background music and wall projections. Figure 2.1. A view of the StudioHome lab. 5

16 One of the current projects using the StudioHome lab concentrates on adapting atmospheres to optimally support people s activities. When people are watching television for example, they might want to have the music turned off, the lights dimmed and the projections on the wall to be static and non-disturbing. Therefore, a system has been developed in which several atmospheres are defined, each having its own settings for light, music and wall projections. The romantic atmosphere for example, covers the room with dim red light and starts playing sweet romantic music. Wall projections are changed accordingly, showing romantic pictures and light animations. The work atmosphere on the other hand is ideal for more serious, official activities. Only bright white light is used and coloured lights, disturbing animations and music are turned off. The current system offers possibilities for people to select the atmosphere that matches their mood or task and even to define their own individual atmospheres. The main problem with the current system is that atmospheres are not switched nor adapted automatically. People have to select a desirable atmosphere themselves and subsequently, they have to use an interactive device, such as a touch-screen, to have the environment changed according to the chosen atmosphere. It would be ideal to have the atmospheres change automatically, based on the knowledge of people s activities. The above application context is in fact a typical smart surroundings project [24]. It aims at investigating, defining and developing a useful infrastructure for ambient systems, that improve people s living conditions. Wireless sensor networks have been used in this context before. For our particular problem, a wireless sensor network has to assist the current atmosphere system in recognising people s activities. We decided to base the activity recognition purely on accurate knowledge of the location of all people present in the room. An alternative solution would have been to use accelerometers [20] for the activity recognition, but this technique requires several sensors to be attached to a person to accurately recognise his activities. Generally, sensors are required around the hips, the thighs, the knees and the wrists in order to distinguish relevant moves. This is expensive in the first place, but it is also annoying for the person carrying the sensors. Another disadvantage of the technique is that it can only provide you with activity information about an individual, but not about a group of people. Even if you know that all people are sitting, you don t know whether they are located close to each other or spread around the room. Consequently the atmosphere system will not be able to choose an adequate atmosphere. Using location information on the other hand, the system will be able to make certain assumptions concerning the activities of people. If the computer is turned on and there is only one person in the room, located close to it, it is very likely that he is using the computer and light and music settings can be changed accordingly. If on the other hand the television is turned on and several people are located on the couches, it is quite certain they are watching television. Similarly, activities like reading, eating or simply walking around can be recognised using location information. 6

17 2.2 Project Objective The goal of our project is to come up with a localisation system that provides finegrained location information about several people present in one room at the same time. The location information should be accurate enough for the researchers at IDE to use it for recognising the people s activities. The emphasis in this report is on the localisation part of the problem rather than on how to apply the location information to activity recognition. First, a study of the existing localisation techniques and algorithms is performed. This study results in an overview of methods that are applicable for WSNs, together with a list of their specific advantages and disadvantages. Second, the scalability of (combinations of) a few techniques is investigated. The question that needs to be answered in the end is: Is one of the localisation techniques that already exist in the literature usable for achieving fine-grained localisation within one room? The specific requirements for the final localisation system, imposed by the researchers at IDE, are listed in the next section. They are important because they restrict the localisation techniques that can actually be used for our project. 2.3 System Requirements The specific requirements for the final system, as stated by IDE, are listed here. The system must: be able to track at least three persons simultaneously; position people using absolute or relative coordinates; at least position people two-dimensionally; guarantee the information to be accurate; be based on small, non-intrusive hardware; be rather inexpensive. 7

18 8

19 Chapter 3 Related Work This chapter presents the prior work that has been done with respect to localisation using wireless sensor networks. A general introduction about localisation in WSNs is given in Section 3.1. In Section 3.2, an overview of existing technologies and algorithms is presented. A table providing a summary of the algorithms presented in this chapter can be found in Section Localisation in WSNs In this section, one can find definitions that are important for understanding the rest of the chapter. Afterwards, common mathematical techniques that are used in many localisation algorithms are presented Definitions Localisation in wireless sensor networks is about knowing the location of any network node at any time. Thereby, nodes can be either mobile, which means that their location can change, or static. The focus of this report is on localising mobile nodes. In this report, it is assumed that sensor nodes only consist of a radio, a processor, memory and a power supply. All additional hardware that might be needed for doing localisation, such as infrared sensors or ultrasound transceivers, is considered extra in this context. Most localisation algorithms assume the presence of a few nodes with prior knowledge of their location: anchor nodes or simply anchors. The position of the other nodes is determined through interaction with or relative to the anchors. From here on, nodes of which the position needs to be determined are referred to as unknown nodes. Trilateration is a common mathematical technique that is used to compute an unknown node s location from the combination of distance estimates of the node to anchors and location information of these anchors. Geometry is applied to deter- 9

20 mine the unknown node s coordinates, as shown in Figure 3.1. A virtual circle with a radius equal to the distance estimate between the anchor and the unknown node is drawn around every anchor. The unknown node is then located at the intersection of all three circles. Figure 3.1. A visual representation of the trilateration algorithm. 3.2 Existing Techniques and Algorithms This section focuses on the techniques and algorithms that are currently available for doing localisation in wireless sensor networks [23]. For all of them, a short description is given, followed by an overview of the main characteristics. The methods presented here form the basis for determining the technique(s) to be used for this specific research project about fine-grained indoor localisation Localisation using GPS Localisation systems for WSNs can be based on the Global Positioning System (GPS), which is a satellite-based localisation infrastructure. At any location on earth, a GPS-receiver can be localised using information of at least four GPSsatellites. The receiver computes the time-of-flight of the different satellite signals as the difference between its local time and the time the signals were sent and converts the times into distance estimates. The receiver also determines the satellites locations from their radio signals and an internal satellite database. From this knowledge, the receiver s position is derived using trilateration, generally with an accuracy of about ten meters. GPS can easily be used in sensor networks, by equipping the sensor nodes with GPS-receivers. Nevertheless, GPS-based localisation in sensor networks has some disadvantages. The first problem is that a GPS-receiver consumes a lot of energy, which is known 10

21 to be a scarce resource on a sensor node. The next problem relates to radio signal propagation in an indoor environment: walls, floors and furniture can disturb or even entirely block the satellite signals. It is often a problem to even detect four satellite signals in an indoor environment. In any case, bad distance estimates result and therefore localisation errors are large indoors. A final disadvantage is the high price for equipping all nodes in a network with expensive GPS-receivers Localisation using Infrared Localisation information in a WSN can also be acquired by equipping the sensor nodes with infrared sensors [27]. Throughout the environment, anchor nodes equipped with infrared receivers are installed. Any unknown node sends an infrared signal at regular intervals. Depending on the sender s location, the signal is detected by a limited number of (different) anchors. Based on this knowledge, the sender s position can be roughly estimated. Room-level granularity is the best accuracy currently obtained with this method. The infrared based solution is suitable for both indoor and outdoor use, but because of the short range of infrared signals, many nodes with receivers are required. This makes the solution quite expensive for large areas. Another disadvantage of the method is the inaccuracy caused by multipath effects and line-of-sight requirements. The first is responsible for false positives: a reflected signal is received instead of a direct one and a receiver incorrectly assumes the sender is within lineof-sight. The second one occurs when there is an object between the sender and the receiver. The sender s signal is not detected, which results in a false negative. Both problems lead to incorrect conclusions about the sender s location Localisation using Sound Sound signals can also be used for localisation purposes in wireless sensor networks [9]. For that, sensor nodes need to be equipped with sound transceivers. In general, ultrasound is used: it is less intrusive since it is not audible for human beings. The first category of algorithms using sound is based on the time-of-arrival or round-trip-time of a sound signal between an unknown node and an anchor. Both methods take a timestamp the moment the sound signal is sent. Depending on the method, the second timestamp is taken the moment the signal arrives at the other node or back at the sending one. The timestamps are used to calculate the sound signal s travel time. The time-of-arrival method requires the nodes in a network to be synchronised, since it uses timestamps taken by different nodes. The distance between the node and the anchor is estimated by dividing the speed of sound by the travel time. Finally the unknown node s location is derived by using for example trilateration. Localisation errors of tens of centimeters can be acquired with this technique. 11

22 An alternative approach using sound is to use time-difference-of-arrival information, where sound signals are combined with radio signals [6, 18]. The principle of the method is as follows: at regular intervals, anchors simultaneously send a radio message and a sound signal. Unknown nodes receive the radio message and somewhat later they detect the sound signal. Based on the knowledge of the speeds of light and sound, a time-difference of arrival between the two signals is computed. From that, a distance estimate to the anchor is derived. The location of the unknown node is then determined using one of the above mathematical techniques, with an accuracy of centimeters. The main disadvantage of the time-of-arrival method is the need for an accurate synchronisation of the sensor nodes. This has proven to be very hard in sensor networks, due to energy constraints and inaccurate processor clocks. Radio messages are needed to synchronise between the different nodes, but they introduce small errors in the time schemes because of the latency inherent to radio communication. The small deviations in the time schemes cause errors in the calculation of the signals travel time, and small errors in the latter cause large localisation errors. A common disadvantage for both the time-of-flight and the time-difference-ofarrival method is that extra hardware is needed. Ultrasound transceivers are still quite expensive and they increase the form factor of a sensor node with at least a factor two Radio-based localisation Localisation in sensor networks can be achieved using knowledge about the radio signal behaviour and the reception characteristics between two different sensor nodes. The quality of a radio signal, i.e. its strength at reception time, is expressed by the radio signal strength indicator (RSSI): the higher the RSSI-value, the better the signal reception. The main advantage of using radio-based localisation techniques is that no additional hardware for the sensor nodes is required. The main disadvantage of the technique is that the measured signal strengths are generally unstable and variable over time, which leads to localisation errors. In this section, two common localisation techniques using radio signal strength information are presented. Afterwards, the proximity idea is discussed, a technique that takes into account the range of radio communication rather than its quality. Finally, a technique for analysing the RSSI behaviour over time is presented. The technique cannot be used for localisation itself, but it can provide useful mobility information about the node to be located. Converting Signal Strength to Distance In theory, there exists an exponential relation between the strength of a signal sent out by a radio and the distance the signal has travelled, as shown in Figure 3.2. In reality, this correlation has proven to be less perfect, but it still exists [3]. 12

23 Signal strength [dbm] a+b*log(x) Distance to sender [m] Figure 3.2. The mathematical relation between signal strength and travelled distance. The above relation forms the basis for the first RSSI-based localisation technique. Anchors broadcast their position at regular intervals. Unknown nodes receive the message and measure the strength of the received signal. This signal strength is converted to a distance estimate, using the exponential relation shown above. Trilateration is used to convert the distance estimate between anchor and unknown node into coordinates for the latter. Localisation errors for this method range from two to three meters at average, with indoor errors being larger than outdoor ones. The main reason for the large errors is that the effective radio-signal propagation properties differ from the perfect theoretical relation that is assumed in the algorithm. Reflections, fading and multipath effects largely influence the effective signal propagation. The distance estimates, which are based on the theoretical relation, are thus inaccurate and lead to high errors in the calculated locations. Fingerprinting Signal Strengths The second method that uses RSSI for localisation is called fingerprinting. This technique is based on the specific behaviour of radio signals in a given environment, including reflections, fading and so on, rather than on the theoretical strength-distance relation. The fingerprinting technique [2, 4, 5, 10, 11, 17] is an anchor-based technique that consists of two separate phases. During the first phase, called the offline phase, a fingerprint database of the environment is constructed. A node is put at a number of predefined points in the deployment area to record the fluctuations in signal strength at these specific points. At each location, the node sends a number of messages and all anchors measure the signal strength of the received messages, or the other way around. The combination of the RSSI-values measured by the different anchors when the node is at a certain location forms the fingerprint of this location: a series of RSSI-values that are representative for that particular location. Per location, a number of fingerprints is stored in a database, needed by the second phase. An example database of a system with four anchors is shown in Figure

24 Figure 3.3. An example grid and corresponding fingerprint database. During the next phase, called the online phase, real-time localisation is performed. An unknown node has to be localised in the deployment area. The unknown node broadcasts a message at regular intervals and the anchors measure the signal strength upon reception of a message. The measured RSSI-values are combined into a RSSI-sample. Afterwards, the best matches between the values in the RSSI-sample and the values stored in the database are searched for. The resulting matches determine the final position of the unknown node. Its location could either be the value of the closest match or an average of a few best matches. The specific algorithm used for matching is not relevant here. The main advantage of using RSSI this way is that the unpredictable RSSI variations in space are handled, which makes the approach a little more accurate. Errors using this method are reduced to an average of one to two meters. The greatest disadvantage of the method is that an offline phase is required for the system to work. The offline phase is in the first place very time consuming. Moreover, the fingerprinting database that is created during the offline phase is location dependent. If one wants to use the same system in another environment or if radical changes to the current environment are made, the offline phase has to be repeated. Proximity-based localisation Proximity-based localisation systems are an anchor-based solution to the localisation problem. These systems derive their location data from connectivity information of the network [15]. Knowledge about whether two devices, i.e. an unknown node and an anchor, in the network are within communication range is transformed into an assumption about their mutual distance and location. The technique is based on the existence of a maximum communication range for a node sending at a given power. Using a proximity-based algorithm, coarse-grained localisation can be achieved. 14

25 The location information can be refined by also measuring the strength of the radio signals between the nodes that are within range of each other [21, 25]. The signal strength can be translated into an estimate of the distance between the two nodes, using for example statistical methods. By combining the location information of the anchors with the distance estimates, the location of an unknown node can be roughly determined. This refinement of the above technique can reduce the errors by 50%. In the literature studied, no real measures are given concerning the average errors attained using proximity-based localisation. It only shows that errors in the range of 50% of the radio range are achievable. A disadvantage of the method is that its performance in a network with high message loss rates will most probably decrease, because in that case, the algorithm can no longer conclude that a specific node is out of the sender s radio range from the fact that the node is not receiving a message from the sender. Analysis of the Radio Signal Strength Behaviour over Time Performing an analysis of the radio signal strength behaviour during a longer period of time, i.e. a few seconds, can provide additional information about the mobility status of a sensor node. Research performed in [12] points out that the variance of the signal strength is much larger when a node is moving than when it is static. Knowledge about a person s mobility pattern does not provide any real location information, but it is useful in combination with other localisation algorithms. In [12], the technique is combined with fingerprinting. The inference of mobility information as well as location information from the radio signal is done using a Hidden Markov Model (HMM). The algorithm in [12] leads to a median localisation error of 1.5 meters and tells whether a node is in motion or not with an accuracy of 87%. 3.3 Summary The table below gives a short overview of the main characteristics of the techniques presented in this chapter. GPS Infrared Ultrasound RSSI applicable not recommended yes yes yes indoors need for yes yes yes no extra hardware cost of high low high n.a. extra hardware size of average average large n.a. extra hardware average ±10 meters ±5 meters ±10 centimeters 1 to 3 meters expected error (room-granularity) 15

26 From the table, we concluded that RSSI can provide us with the cheapest localisation system possible, while the form factor of the sensor nodes is not increased. The technique is applicable indoors and the errors achieved with a RSSI-based system seem to be promising compared to the more expensive systems: only sound-based systems perform better. 16

27 Chapter 4 A Preliminary Study of Applicable Techniques In this chapter, the preliminary research towards the applicability of the existing localisation techniques and algorithms for this specific project is described. In Section 4.1, the available hardware and software are described. In Section 4.2, the localisation technique that is eventually chosen is presented and this choice is motivated. Experimental results that justify the usability of the technique are presented in Section 4.3 and in Section 4.4, one can find a short introduction to the chosen localisation systems. 4.1 Available Hardware and Software At the department of Computer Science, several types of sensor nodes are available: EYES nodes, MICA2 nodes, MICA2DOTs and TNOdes (see Figure 4.1). (a) EYES node (b) MICA2 node (c) MICA2DOT (d) TNOde Figure 4.1. Four different sensor nodes. We decided to use the TNOdes for our project, because they are currently the smallest, cheapest and newest nodes. TNOdes have a Chipcon CC1000 radio transceiver, an atmel ATmega128L processor and a 4Mbit atmel dataflash AT45DB041B EEP- ROM memory. Equipping the TNOdes with extra sensors is possible using the node s available plug-in connectors. 17

28 A wide variety of operating systems for sensor nodes exists, such as TinyOS, Mionos and Mantis OS. Of all known operating systems, TinyOS [1] is the most popular one. It is an open-source operating system initially developed at the University of California, at Berkeley. TinyOS has a component-based architecture and components providing different functionalities, such as timers, data storage in the EEPROM and radio communication are already included in the operating system. Additionally, programmers can write their own components and connect them to existing ones to create a working program. The execution is event-driven. This means that interrupts are generated whenever certain events occur, for example at the expiration of a timer. Subsequently, an event handler, which is in fact a function, is called to handle the event adequately. TinyOS is currently available for many hardware platforms, including the TNOdes, and it is easily extensible and well-tested. Therefore, we decided to use TinyOS for this project. Power consumption issues for sensor nodes have gained an increasing amount of attention during the last few years. Every day, new efforts are made to reduce the power requirements of sensor nodes. This is also reflected in the existing MACprotocols for TinyOS. B-MAC [8], S-MAC [26] and T-MAC [22] are developed almost simultaneously by different institutes and they show a different amount of effort in reducing the energy consumption of a sensor node. In B-MAC, these efforts are minimal. The radio and processor for example are turned on all the time, unnecessarily consuming precious energy. S-MAC is the first MAC-protocol to make attempts on saving energy, mainly by providing duty-cycling. This means that all nodes in a network are synchronised and each node has its own scheme for sending and receiving messages. Outside of send and receive intervals, the node s radio is put to sleep because no messages can be expected anyway. This way a large amount of energy is saved, since the radio is one of the most powerconsuming components of a sensor node. T-MAC is the most recent MAC-layer, derived from S-MAC. It adopts the same ideas of duty-cycling, but the duration of the wake and sleep intervals is adapted depending on the message activity in the sensor network. This means that the node s radio can be turned off again if there are no messages to be sent or received during the wake interval, in order to save even more energy. Because we want our localisation system to run for the longest period possible, without having to change the batteries of the TNOdes, we decided to use T-MAC. 4.2 Chosen Localisation Technique We eventually decided to design and implement a RSSI-based system to solve the localisation problem of this Master s project. The main reason for this is that it can be used without having to equip the TNOdes with additional hardware. This way, two of the initial system requirements (see Section 2.3) are already fulfilled: 18

29 1. The small form factor of the sensor nodes is guaranteed, enabling the localisation to be non-intrusive for users; 2. The nodes are as cheap as possible, because no extra sensors have to be bought, installed, or designed. Moreover, prior experiments with RSSI-based localisation algorithms have shown that RSSI is indeed usable for performing indoor localisation (see Chapter 3). Experiments were mostly performed in areas of tens to hundreds of square meters and led to average localisation errors of a few meters (see Section 3.2.4). The objective of this Master s project is to try to apply the existing localisation algorithms on a smaller scale. Thereby, it is assumed that the error will scale down accordingly. 4.3 Indoor Characteristics of Radio Signal Strength Before implementing a localisation system, we conducted some experiments to investigate the actual behaviour of radio signals in a small indoor environment (i.e. one room). The experimental results would either confirm or refute the assumption that RSSI is usable for localisation within one room. Moreover, they can provide insight into which algorithms will most probably give the best results. All the conducted experiments use the same setup, which consists of one sending node and one receiving node. The receiving node is attached to a computer for data transmission purposes and its location is kept fixed during all the experiments. The sender is put either at a fixed location or moved along a straight line with steps of 50 centimeters. When the sender is turned on, it starts sending messages containing its identifier and a sequence number every 200 milliseconds. Upon reception of a message, the receiver measures its signal strength. Afterwards, it forwards the data to the computer. This process continues until 1,000 messages have been received. Using this approach, the three following properties of radio signal strength are investigated: 1. The reproducibility of radio signal strengths; 2. The relation between signal strength and distance on a short range (meters); 3. The influence of dynamic factors, such as moving people or moved objects, in a node s environment on the signal strength. More details about the experiments, the experimental results and the conclusions are presented in the next sections. 19

30 4.3.1 The Reproducibility of RSSI measurements First of all, it is important to know whether the results of RSSI measurements are reproducible. This is mainly important for the fingerprinting technique of Section 3.2.4, which relies on the reproducibility of RSSI measurements. In order to investigate the reproducibility, three series of measurements are performed. For the first series, both the sender and the receiver are placed at a fixed location. The sender is turned on and the receiver measures and stores the RSSI values of 1,000 messages. The series is repeated twice: the first time using exactly the same sender and receiver; the second time the original sender is replaced by a different sensor node of the same type, which is executing the same sender code. If the results of RSSI measurements are reproducible, all series of measurements should produce the same or at least very similar results. Figure 4.2 shows the results of the experiment first time second time first time second time RSSI [dbm] RSSI [dbm] Sequence number Sequence number (a) Same TNOde (b) Different TNOde Figure 4.2. Reproducibility of RSSI values. The same signal strengths are measured if the experiment is repeated using exactly the same hardware components, which is shown in Figure 4.2(a). Slight variations in signal strength are shown in Figure 4.2(b). These are due to the use of a different hardware component for the sender, i.e. another sensor node of the same type, running the same sender code. The signal range of an identical component is the same, but the average value deviates. This can nevertheless be corrected by calibrating the sensor nodes Relation between RSSI and Distance A second experiment investigates the relation between radio signal strength and distance. One wants to know whether the theoretical exponentially decreasing relationship between both factors is in any way present in a small indoor environment. In this experiment, the receiver is placed at a fixed location. Between every two series of 1,000 RSSI measurements, the sender is moved along a straight line of four meters, with steps of 50 centimeters. This way, the sender is subsequently put 20

31 at an increasing distance from the receiver. The 1,000 RSSI values resulting from one series of measurements are averaged to get an estimate of the expected RSSI at that specific distance. The results of this experiment are plotted in Figure 4.3. The plot indicates that there is no perfect exponentially decreasing relation present. An overall decrease of the signal strength with increasing distance can be seen for the range of the experiment. It is nevertheless not sure whether this behaviour continuous after five meters RSSI [dbm] Distance to sender [cm] Figure 4.3. The actual RSSI behaviour indoors Influences of a Dynamic Environment The last experiment looks at the influence of dynamic factors in the environment on a location s signal strength. First, one wants to know the consequences of changing the orientation of a node on the measured signal strength. Second, the signal changes caused by moving objects in the environment are investigated. Third, the influences of people walking around close to the nodes have to be considered. The plots in Figure 4.4 show the influences of these three factors. Figures 4.4(a) and 4.4(b) show the signal strength values before and after changing the node s orientation 180 degrees and before and after moving some pieces of furniture, respectively. In Figure 4.4(c) the signal strength values with and without people wandering around can be seen. It is clear that slight changes in the position s of objects has the least influence. Changing the node s orientation causes a shift in the signal strength but not really an increase in signal variance. Moving people on the other hand increase the signal s variance but only slightly change the average signal strength Conclusions The experiments show that indoor radio propagation differs from the ideal situation. Only an overall decreasing trend in radio signal strength over a distance is noticeable. The RSSI behaviour is reproducible and it suffers moderate but welldefined variations as a consequence of dynamic factors. Therefore, it can be con- 21

32 original orientation changed orientation before moving objects after moving objects RSSI [dbm] RSSI [dbm] Sequence number Sequence number (a) Changing node orientation (b) Moved objects without moving people with moving people average with moving people RSSI [dbm] Sequence number (c) Moving people Figure 4.4. RSSI changes due to dynamic factors. cluded that RSSI is applicable for indoor localisation, if only these problems are kept in mind and handled adequately. 4.4 Two Different Localisation Systems Given the conclusions of the previous section, two localisation algorithms remain an option: fingerprinting and proximity-based localisation. These techniques were introduced in Section Both techniques are anchor-based and do not rely on a perfect relation between radio signal strength and distance. The proximitybased solution implemented for our problem will be extended with functionality for signal analysis, as is presented at the end of Section Detailed information about the final systems can be found in Chapters 5 and 6, respectively. 22

33 Chapter 5 The Fingerprinting System In this chapter, details about the fingerprinting algorithm as it was implemented for this project can be found. In Section 5.1, a motivation for implementing this algorithm is given. Specific setup and design details are presented in Section 5.2. Details about the implementation follow in Section 5.3. In Section 5.4, the results of experiments that were conducted to test the system are presented. The applicability of the system for this specific project is reviewed in Section Motivation The main reason for implementing the fingerprinting algorithm is its quite good performance on a larger scale. Based on conclusions of former implementations found in the literature (see Section 3.2.4), it seems to be a fair and reasonable assumption to expect that localisation errors will scale down if the localisation area does so, while the number of anchors is kept the same. In any case, one can be sure that fingerprinting will perform better than any algorithm that purely relies on transforming RSSI measurements into distances, as was explained in Section A second important reason is that, if the algorithm works, it provides the researchers at IDE with detailed location information, namely two-dimensional absolute coordinates. From these coordinates, accurate information about the objects at which the people are located can be derived. The coordinates can tell the researchers whether a person is really at a location and using an object, e.g. a couch, or only close to it but not using it. 5.2 System Setup and Design The idea of the fingerprinting algorithm designed and implemented for this Master s project is the same as explained in Section The algorithm is based on the uniqueness and reproducibility of RSSI values at specific locations and on the assumption that there exists a piece wise linear fit for the effective relation between signal strength and distance. The signal strength distance relation does not 23

34 need to be perfectly decreasing, because unexpected variations are automatically handled by the algorithm. In this section one can find general information about the approach followed for the fingerprinting system. Details about the specific algorithms and programs used are given in the next section. The first step towards the final system is to set up a localisation infrastructure. It consists of four anchors that are installed at fixed locations in a room of 7 by 4 meters. Different setups are tried: Anchors are installed in the middle of the walls, against the ceiling (see Figure 5.1(a)); The ceiling area is virtually split up in four rectangles and an anchor is placed in the middle of each rectangle (see Figure 5.1(b)); Anchors are put in the middle of the walls, at shoulder-height (see Figure 5.1(c)). (a) Setup 1 (b) Setup 2 (c) Setup 3 Figure 5.1. The different anchor setups used. All anchors in the system are connected to a personal computer using a USB connection. In fact, only one USB connection is necessary to enable the data transfer between the sensor network and the computer. The rest of the network is then wireless and depends on batteries for power supply. We decided to connect all anchors to the computer using USB connections, in order to take away their energy restrictions. After setting up the infrastructure, the fingerprinting algorithm can begin. It starts with an offline phase, during which a virtual two-dimensional grid is laid out in the room at table height. The intersection of any two grid lines is called a grid point. A sensor node is put on all grid points in the room and subsequently, the node broadcasts messages for a while and all anchors measure the received signal strength for each message. The set of RSSI values that are concurrently measured by the different anchors forms a fingerprint: a combination of RSSI values that are to be expected when the sender is put at that specific location again. Per grid point, 24