Advanced Indoor Localisation Based on the Viterbi Algorithm and Semantic Data Jens Trogh * , David Plets * , Luc Martens * and Wout Joseph * * Department of Information Technology, Ghent University/iMinds, Belgium, jens.trogh@intec.ugent.be Abstract—In this work a real-time indoor localisation system based on the Viterbi algorithm is developed. This Viterbi prin- ciple is used in combination with semantic data to improve the accuracy: i.e., the environment of the object that is being tracked and an adjustable maximum speed. The developed algorithm was verified by simulations and with experiments in a building- wide testbed for sensor and WiFi experiments. Compared to a reference algorithm without Viterbi or semantic data, the results indicated a significant improvement: the mean accuracy and standard deviation improved by respectively 26.4% and 63.9%. Index Terms—Localisation, Viterbi Algorithm, Semantic Data, Wireless Networks, Indoor Environment I. I NTRODUCTION Indoor localisation systems have applications in many do- mains, think of the healthcare sector, agricultural sector, indus- trial sector, cultural sector, etc. Examples of these applications are: tracking of elderly, monitoring of animals, equipment tracking and museum guidance. To locate an object, most lo- calisation systems use a mobile node and a fixed infrastructure, which consists of static nodes. These static nodes are con- nected and form the wireless network. The mobile and fixed nodes exchange signals and the characteristics of these signals are used to estimate the position of the mobile node. Due to the complexity of many indoor environments, localisation systems are often not sufficiently accurate. Current state-of- the-art localisation systems try to improve the accuracy by using new technologies like Ultra Wideband (UWB) [1]. The very large bandwidth enables highly accurate localisation but the typically smaller ranging coverage makes it more suited for short-range applications. Other state-of-the-art systems rely on user interaction or route prior knowledge but this is not always wanted or even possible [2]. In this paper, an advanced indoor localisation algorithm for tracking a person in real- time through a building, is presented. To avoid expensive hardware costs or a time-consuming measurement campaign, the existing WiFi or ZigBee infrastructure and an advanced network planner are used. II. METHODOLOGY A. Localisation algorithm The localisation algorithm is based on the Viterbi algo- rithm [3]. This dynamic programming algorithm is used to determine the most likely sequence of hidden states, called the Viterbi path, resulting in the sequence of observed events. To apply this technique on a localisation algorithm, the states have to be interpreted as real locations on a floor plan. Then, this principle comes down to determining the most likely sequence of positions instead of only the most likely current position. All possible trajectories are kept in memory and each trajectory has an associated cost. This cost is the sum of Mean Square Errors (MSE) between measurements and reference values (see Section II-C) and is used as decision metric. To apply the Viterbi principle in a useful manner (improve the accuracy), we have to restrict the number of allowed transitions between two consecutive locations. Therefore, we use semantic data: the environment of the object that is being tracked and an adjustable maximum speed. In this way it is assured that no walls are crossed (doors are used to leave a room) and no unrealistically large distances are crossed within a given time frame. Overall, this leads to realistic and physically possible trajectories. To the best of the author’s knowledge this is the first localisation algorithm that uses this combination of techniques. B. Start position Because the most likely sequence of positions is determined, the developed localisation algorithm is sensitive to a wrong start position. One could start off in the wrong room, which implies a certain recovery time before predictions can be accurate again, because walls cannot be crossed. To counteract this, additional start positions are added as soon as the tracking begins. These additional start positions lie on circles around the best initial prediction. In this way the algorithm can easily correct itself by switching to another trajectory when new measurements suggest being located inside a different room. By using this technique also the previous positions will be set right. C. RSSI fingerprinting using heuristic indoor network planner The developed localisation algorithm relies on a Received Signal Strength Indicator (RSSI) fingerprinting technique to estimate the most likely position by comparing the measure- ments with reference values from a radio map. This radio map (also known as fingerprint database) contains the path losses to all fixed nodes, for each possible position (grid point) on the floor plan where the localisation takes place. The size of the fingerprint database will depend on the size of the floor plan, the resolution of the possible positions (grid size) and the number of access points (APs). The path losses can be calculated with a theoretical model or obtained via a measurement campaign. Because the latter is an expensive