Activity classification and dead reckoning for pedestrian navigation with wearable sensors

IOP PUBLISHING Meas. Sci. Technol. 20 (2009) 015203 (10pp) MEASUREMENT SCIENCE AND TECHNOLOGY doi:10.1088/0957-0233/20/1/015203 Activity classification and dead reckoning for pedestrian navigation with wearable sensors Zuolei Sun 1, Xuchu Mao 1,WeifengTian 1 and Xiangfen Zhang 2 1 Navigation and Control Lab, Department of Instrumentation Engineering, Shanghai Jiao Tong University, Shanghai 200240, People s Republic of China 2 College of Mechanical and Electronic Engineering, Shanghai Normal University, Shanghai 201418, People s Republic of China E-mail: sunzuolei@gmail.com Received 29 January 2008, in final form 24 September 2008 Published 13 November 2008 Online at stacks.iop.org/mst/20/015203 Abstract This paper addresses an approach which integrates activity classification and dead reckoning techniques in step-based pedestrian navigation. In the proposed method, the pedestrian is equipped with a prototype wearable sensor module to record accelerations and determine the headings while walking. To improve the step detection accuracy, different types of activities are classified according to extracted features by means of a probabilistic neural network (PNN). The vertical acceleration data, which indicate the periodic vibration during gait cycle are filtered through a wavelet transform before being used to count the steps and assess the step length from which the distance traveled is estimated. By coupling the distance with the azimuth, navigation through pedestrian dead reckoning is implemented. This research provides a possible seamless pedestrian navigation solution which can be applied to a wide range of areas where the global navigation satellite system (GNSS) signal remains vulnerable. Results of two experiments in this paper reveal that the proposed approach is effective in reducing navigation errors and improving accuracy. Keywords: wearable sensors, pedestrian navigation, activity classification, dead reckoning (Some figures in this article are in colour only in the electronic version) 1. Introduction Pedestrian tracking and navigation are now being widely researched and applied. A large part of pedestrian travel takes place indoors or in light indoor environments where the global navigation satellite system (GNSS) cannot provide uninterrupted and reliable position information due to the low signal-to-noise ratios, poor satellite geometric distribution and multipath effect. As such, self-contained approaches with wearable sensors based on the dead reckoning (DR) principle which does not require any outside infrastructure support are preferable in pedestrian navigation. Wearable sensors refer to a family of sensors which are generally small sized, and can be easily mounted on the body. These include inertial sensors such as gyroscopes and accelerometers, magnetic sensors such as magnetometers and electronic magnetic compasses as well as various other types of sensors such as thermometers, barometers and cameras. These wearable sensors are often integrated into the pedestrian navigation module (PNM) to provide the information for calculating the position. With a known starting point, the DR approach provides a two-dimensional relative positioning solution through measurements of azimuth, velocity and time. In a conventional DR system, the distance traveled is usually measured by a vehicle-carried odometer. However, in the case of the pedestrian dead reckoning (PDR) problem, movement is carried out by feet, and so generally the distance traveled is estimated by multiplying the number of steps by the step length. In most cases, the step count is calculated using pedometers and accelerometers. In [1], conventional sports pedometers and accelerometers were used to count steps 0957-0233/09/015203+10$30.00 1 2009 IOP Publishing Ltd Printed in the UK

while two- or three-axis compasses and rate gyroscopes were employed to measure azimuths. This paper compared different wearable sensors for PDR in four field tests which were part of the Equator project. The step length, which is a key component of the PDR algorithm, is a time-varying process. It is strongly correlated to several other characteristics of the gait, especially the velocity and step frequency of the pedestrian. In some studies, fixed average step lengths were adopted, as in [1, 2], while in others, the step length was assessed in real time using physiological models. In [3], Levi et al suggested that the step length could be estimated online based on a linear relationship between the measured step frequency and step length. This method was also used by Lee et al in [4]. Similarly, the linear relationship between walking speed and step length was used in the PNM developed for Leica Vectronix AG by Ladetto et al from Swiss Federal Institute of Technology [5 8]. In this paper, a wearable sensor module with a built-in accelerometer and a magnetic compass was mounted on the test subject to measure the accelerations and the azimuths. The step length was evaluated using Levi s model to establish its relation with the step frequency which was acquired by analyzing the vertical acceleration data. Experiments were carried out to tune the step length model. Another key element of the PDR algorithm is the step detection which serves the step count. Several common methods include: comparing the acceleration values with the predefined thresholds and taking the minimum step period into account [3, 9, 10], identifying the step occurrences using pitch signals [2] and comparing the acceleration variances with a threshold [5]. However, peaks of acceleration also occur in irregular motions such as jumping or avoiding obstacles on a crowded road. These peaks and those that occur during regular walking have very similar peak values and time intervals between successive peaks. If all the acceleration data are fed into the peak detection algorithm, step misdetections will occur. So to make the step count more reliable and to further improve the accuracy of the distance traveled calculation, a probabilistic neural network (PNN) is introduced to the PDR algorithm to identify only the valid acceleration readings acquired resulting from regular walking movements. This activity classification method, principally derived from the fields of context awareness and wearable computing, is validated by primary experimental results. A very successful work in this area was described in [11], where five biaxial accelerometers were attached to different parts of the test subjects and by decision tree classifiers, an accuracy rate of 84% was achieved in recognizing everyday activities. This paper is structured in the same order as the various processing stages of our proposed pedestrian navigation solution as shown in figure 1. Section 2 introduces the wearable sensors and data collection methods, section 3 addresses activity classification with feature extraction and the PNN classifier, section 4 discusses the PDR algorithm in detail and in section 5, experimental results are presented to demonstrate the effectiveness of the proposed algorithm. Accelerations Wearable sensors Feature extraction PNN activity classifier Activity classification Wavelet transform Azimuths Step detection Step length assessment Position calculation Pedestrian dead reckoning Moving average filter Figure 1. Diagram of the proposed pedestrian navigation algorithm. 2. Data collection 2.1. Wearable sensor In this study, the accelerations and azimuths of the pedestrian were collected with a wearable sensor module, which is embedded with a micro-electro-mechanical systems (MEMS) chip AK8976A marketed by Asahikasei Microsystems Co., Ltd (AKM) hybridizing a tri-axial accelerometer and an electronic magnetic compass. The built-in accelerometer has a typical accuracy of ±22 mg; the built-in electronic magnetic compass has an offset magnetic field compensation range of ±2.0 mt. By processing these data with an external CPU, Microchip PIC16LF877, azimuth data can be obtained. The sampling frequency is 64 Hz, which is more than sufficient compared to the 16 Hz frequency required to recognize the activities of the pedestrian [11]. The sensor module was mounted on the pedestrian s belt. The signals from the builtin accelerometer and compass were transmitted to a laptop wirelessly over bluetooth. The outputs of the sensor module were annotated with time indicated by the laptop. Figure 2 shows the wearable sensor module and how it was attached to the test subject. Figure 3 shows the typical vertical (the y-axis) acceleration data acquired when the test subject is walking straight ahead, standing still and making irregular movements such as jumping or avoiding obstacles on a crowded road. In practice, completely motionless stops of a pedestrian rarely happen. For instance, when stopped by a traffic light, although the pedestrian is not walking, he is not completely static either. Trials have shown that they frequently shuffle and such movements are treated as irregular motions [9]. 2.2. GPS module The positioning results yielded from the Superstar highsensitive GPS receiver were compared to the PDR field test 2

AK8976A Y X Z Figure 2. The wearable sensor module and the way it was attached to the test subject. Walking Standing still Irregular motions Step misdetection may occur Vertical acceleration Vertical acceleration Walking Irregular motion 0 5 10 15 20 25 30 35 40 45 50 Figure 3. Raw vertical acceleration data under different movements. results. The GPS data were also annotated with time indicated by the laptop, and a nearest-neighbor algorithm was applied to ensure the synchronization between the GPS and the wearable sensors. 3. Activity classification In the proposed pedestrian navigation algorithm, like in many other studies, the step count is determined using a peak detection algorithm, based on the acceleration value threshold and minimum step period. However, as is shown in figure 3, the peaks of vertical acceleration occur not only in walking, but also in irregular motions. Figure 4 shows the overlapped waveforms of the vertical acceleration both in walking and irregular motions, each over a period of 7 s. The green circle highlights that the two waveforms have very similar peak value and time interval between successive peaks; therefore, if the peak detection algorithm was applied to the irregular 0 1 2 3 4 5 6 7 Figure 4. Waveforms of the vertical acceleration in walking and irregular motions, each over a period of 7 s. accelerations directly, false steps would be detected. With the false steps adding to the total distance traveled, the alongtrack error would increase as a result. Therefore, the activity classification is implemented before the PDR algorithm to avoid step misdetection. In this study, the pedestrian activities are classified by means of a probabilistic neural network. 3.1. Feature extraction If the raw accelerometer data are used directly as inputs to the classifiers, the activity classification frequently produces poor results. However, it is possible to use small preprocessing routines to obtain some appropriate features, which enhances the quality of classification [11, 12]. In this study, features are extracted from the raw accelerometer signals via a sliding window of 256 samples, 128 of which overlap with consecutive ones. Feature extraction on sliding windows with 50% overlap 3

50 45 40 35 Standard Deviation 500 450 400 350 Energy 5.8 5.6 5.4 Entropy Walking Irregular motions Standing still 30 300 5.2 25 250 5 20 200 4.8 15 10 150 100 4.6 5 50 4.4 0 0 50 100 150 200 0 0 50 100 150 4.2 200 0 50 100 150 200 Figure 5. The features extracted from different activities. has been explained in previous works such as [11]. In our experiment, with an acceleration sampling frequency of 64 Hz, each sliding window covers a time interval of 4 s. The window of 4 s is used to sufficiently capture cycles of the pedestrian activities; moreover, the size of the 256 sample window enables fast computation of fast Fourier transforms (FFTs). This approach was employed in [11] to successfully extract a wide variety of features. However, since in the proposed algorithm the objective of activity classification is to select useful acceleration samples for step detection and step length assessment, it is unnecessary to recognize all types of activities with high accuracy, and to distinguish between walking, stop and irregular motions is enough. Therefore, the feature vector in this work contains three elements: Standard deviation. The range of possible acceleration data varies considerably between the different activities and as such it is an important feature for the PNN when attempting to distinguish between the three possible motions. Energy. The energy feature is widely considered in activity measurement and recognition, [11]. It is defined as E = L i=1 f i, in which f L i is the component produced by FFT, and L represents the length of the sliding window. In this paper, a window length of 256 is considered. The energy is the normalized sum of the discrete FFT component magnitudes of the acceleration samples. The dc component (mean of the samples) should be excluded from the energy computation. Additionally, the sum of the FFT component instead of the squared FFT component should be adopted, unlike usual cases as [11], since it is found that sharp deterioration in the recognition accuracy would occur when the sum of the squared FFT component was used. Frequency-domain entropy. The entropy of a probability distribution is simply the expected value of the information of certain distribution, and it represents the average amount of information obtained from each sample of a data stream [13]. Here, the frequency-domain entropy of acceleration signals is formulated as follows: L H frequency-domain = P i log(1/p i ), (1) with P i = i=1 f i L i=1 f i, (2) where P j denotes the probability of each discrete FFT component of the acceleration signals. The dc component of the FFT should also be excluded from this calculation. This aids discrimination between activities with similar energy values [11]. The features extracted from three classes of activities including walking, standing still and irregular motions are shown in figure 5. 3.2. Probabilistic neural network (PNN) Artificial neural networks, which gained prominence in the area of pattern recognition, have several properties that make them attractive for human activity classification. In this study, probabilistic neural network is employed as the activity classifier, because of its relatively simple implementation, robustness to noise and its sound statistical foundation in Bayesian estimation theory. Most importantly, it can be used in real time since its training is easy and instantaneous [14, 15]. The PNN was first proposed by Specht in [14]. The architecture of a classical PNN is shown in figure 6. The input layer does not perform any function of computation and simply distributes the input to the pattern layer. In the pattern layer, there is one node for each training example. Each pattern node 4

x INPUT LAYER X 11 X 1i X a1 ϕ11 m ϕ1i ϕa 1 X ai ϕai PATTERN LAYER Vertical acceleration Σ Σ SUMMATION LAYER f 1 Cx ( ) Figure 6. Architecture of the PNN. forms a product of the input pattern vector x: ϕ ai = 1 (2π) d/2 σ d exp f a DECISION LAYER [ (x x ai) T (x x ai ) 2σ 2 ], (3) where x ai is the ith training vector from class a, d denotes the dimension of the input feature vector of the test pattern, and σ is the smoothing parameter. The summation layer neurons calculate the maximum likelihood of pattern vector x being classified into class a by summarizing and averaging the inputs from the pattern layer neuron node that belong to the same class: f a = 1 m ϕ ai, (4) m i=1 where m denotes the number of training vectors in class a. The decision layer neurons are three inputs and produce one output, deciding which class the input vectors belong to. Assuming that the aprioriprobabilities for each class as well as the misclassification losses are the same, the classification of pattern vector x is made in accordance with the Bayes strategy based on the output of the summation layer: c(x) = arg max{f a (x)}, a = 1, 2,...,n (5) where c (x) represents the estimated class of pattern vector x, and n is the number of classes in the training samples. The 3D feature vectors mentioned in section 3.1 serve as the inputs for the PNN, and the outputs indicate whether the pedestrian is walking or not. Thirty nine groups of samples, 13 for each class of activities, were used to train the PNN, and 195 groups of samples were used to test its performance. The activity classification results are as follows: the accuracy rate of identifying walking, standing still and irregular motions were 98.5%, 100% and 83.1%, respectively. While the activities of standing still and walking were almost always correctly classified for each sample, irregular motions suffered from a reduction in accuracy due to the sporadic nature of the sensor data associated with this activity. However, since the objective of the activity classification in this study is to classify Raw signals Filtered signals 0 1 2 3 4 5 6 7 8 Figure 7. Raw and smoothed vertical acceleration acquired in walking. the acceleration samples which represent walking rather than special context awareness researches such as [11, 12], the recognition accuracy of walking is the first priority. The experimental results of the pedestrian navigation algorithm with and without activity classification will be presented in section 5. 4. Pedestrian dead reckoning (PDR) Dead reckoning is the process of estimating the present position by projecting azimuth and speed from a known starting point. In vehicular navigation application, the traditional DR system and strapdown inertial navigation system (SINS) are always utilized. However, taking the cost and accuracy into consideration, neither is suitable for pedestrian navigation [3, 8]. Therefore, a PDR algorithm tailor made for pedestrian navigation is proposed, in which the accelerometer is for different uses than in the classical DR system and SINS. The approach is to analyze the accelerometer data to detect steps, count step number and assess the step length, which is integrated with the azimuths measured by a compass to calculate the position. 4.1. Acceleration data reprocessing All three acceleration components during walking exhibit a cyclical pattern. As shown in figure 3, they-axis acceleration data provide the strongest indication of gait events; however, the waveform of the raw signals from accelerometers is still a bit disordered as in figure 7, since human motions are not absolutely regular and some noise is introduced while walking. To improve the performance of step detection (see section 4.2), it is better to smooth the raw acceleration signals before the subsequent processing [8, 16]. The wavelet transform (WT), an extension to the Fourier transform, is a widely-used tool of data smoothing. It projects the original signal to wavelet basis functions and provides a mapping from the time domain to the timescale plane. The 5

wavelet functions localized in the time and frequency domain are obtained from a single prototype wavelet called mother wavelet. By applying scaling and translation operations to the mother wavelet, a family of wavelet functions is created with the same shape as the mother wavelet but of different sizes and locations. This signal decomposition breaks down the original signal into child signals with different frequencies thus enabling the raw sensor data containing both true signals and high-frequency noise to be smoothed. In this paper, the Daubechies 4 (DB4) wavelet base is introduced to remove the outliers or major noise spikes. It is found from experiments that when the maximal decomposition scale was 3, the ideal result could be obtained. Consequently, the raw signals were decomposed and the wavelet coefficients of every scale were obtained. Since the signals in the third scale contain the step frequency of most test subjects, the wavelet coefficients of the third scale served as the input of the step detection algorithm. The result after smoothing is shown in figure 7. 4.2. Step length assessment The step length assessment is fundamental to the calculation of distance traveled. In this paper, two points are crucial to the proposed step length assessment algorithm: (1) Estimating the step frequency and (2) Determining the relationship between the step length and step frequency. The method most often used for frequency analysis is FFTs. However, as discussed in section 3.1, the activity classification algorithm evaluates 256 acceleration samples, which is too few for an FFT to obtain an accurate step frequency resolution. Therefore, the time interval between consecutive peaks of y-axis accelerations indicating the step cycle is employed instead to determine the step frequency. A peak detection algorithm using another sliding window (differing from the windowing function applied in section 3.1) is employed to find the positive peaks of accelerations [3, 4]. Since an inappropriate length of the sliding window will lead to misdetection of peaks, it is pre-determined by the approximate step frequency acquired from an FFT calculation. In bio-mechanics, the cycle of human gait is generally considered as an interval of time during which one sequence of a regularly recurring succession of events takes place in the person s walk. The regular walking motion is accompanied by periodic variation in vertical accelerations. When human feet move faster, for stability reasons, they usually stretch out their legs further and, as a result, the step length increases, [17]. In other words, if the pedestrian walks faster, both the step length and step frequency increases. Furthermore, in [3], Levi et al argued that the step length could be estimated based on a linear relationship between the measured step frequency and step length. This relationship is also used by Lee et al to assess the step length [4]. Levi models the step length as follows: S = S 0 + m(f f 0 ), (6) where S denotes the step length, f denotes the corresponding step frequency, S 0 is the default step size, f 0 is the step Step frequency (Hz) 3 2.8 2.6 2.4 2.2 2 1.8 1.6 1.4 1.2 Step frequency Step length 1 0 50 100 150 200 0.2 Sample number Figure 8. Step frequency and step length samples in the experiments. frequency corresponding to S 0, and m is the slope of the model curve. S 0, f 0 and m are constants, which are particular to the person with different physical features. Levi s model is adopted in this paper due to its easy implementation. To determine the parameters, five test subjects were employed in the experiments. Each subject was asked to cover a fixed distance of 25 m over 40 separate trials, 10 times at a fast walking pace, 20 times at normal speed and a final 10 times walking relatively slowly. For each test, the time taken and number of the total steps were recorded. The average step length and step frequency could then be determined. The 200 step frequency entries and corresponding step length samples obtained from the experiments are shown in figure 8. Finally, a linear curve fitting algorithm was applied to the samples and Levi s model can be simplified as a linear equation: 1 0.8 0.6 0.4 Step length (m) S = K 1 f + K 0. (7) The experimental results reveal that the model with K 1 = 0.4504 and K 0 = 0.1656 did well in step length assessment, when the step frequency is between the interval of (1.35, 2.45). However, Levi s model depends on the walking features of individual pedestrians, and its parameters may vary for different groups of subjects. Consequently, the parameters used in this study may not be appropriate for other groups of subjects with different physical features including age, stature, gait, etc. The most effective method to improve the precision of step length assessment is to calibrate the parameters for each subject. For example, if the subject walked at a steady pace from the starting point to the end point, and an accurate position of both can be obtained from the digital map, then the parameters could be calculated based on the distance obtained from the map. The step number [3] or step lengths could be updated online using the data fusion method combining with GPS whenever a signal is available [2]. This will be the focus of future work. The proposed step length assessment algorithm is outlined in algorithm 1. 6

Algorithm 1: 1. Calculate FFT over 256 y-axis acceleration samples following smoothing via WT (see section 4.1). 2. Find the dominant frequency component f d (an approximation of the step frequency), and obtain the length L 2 of the sliding window using L 2 = fs f d,wheref s is the sampling frequency (64 in this instance). This step ensures that the length of the sliding window is of suitable size for the peak detection; otherwise, false peaks will be captured, such as the small peak between two large peaks. 3. Shift the sliding window, the oldest acceleration sample is eliminated while the new sample is introduced. The middle sample in the window is compared to all the others (if L 2 is an even number, the middle one can be approximate). If it satisfies the following two conditions, a potential step is detected, or else, another sample is taken. It should be greater than all other samples. The elapsed time since the last potential peak should be larger than the minimum step period, which is 1 3f d here. 4. Count the number of samples between this peak and the last peak, and divide it with the sampling frequency, the result is considered as the period of this step, so the step frequency which is the reciprocal of the step period can be obtained. 5. The step length can then be estimated by substituting the step frequency into equation (7). A step period Figure 9 shows the result of step detection in 8 s, the interval between two red cycles denotes the step period, on which the step length assessment is based. Vertical acceleration 0 1 2 3 4 5 6 7 8 Figure 9. Step detection in 7 s. 4.3. Position calculation A moving average filter is applied to the electronic compass readings to reduce the effect of drift arising from magnetic interference. Discontinuities in the filter s output occur if the input data are represented in degrees (as shown in figure 10). As such, the sine and cosine of the raw azimuths are used as the inputs to the filter and subsequently in calculating the position of the pedestrian. Then the latitude and longitude can be calculated using the following equations with a given starting point [18]: ϕ t+1 = ϕ t + S cos A, (8) R m Figure 10. Moving average filter ran on raw data and the value of sine of compass measurements. 7

with λ t+1 = λ t + (D t+1 D t ) sin A cos A, (9) ( ( π D t+1 = Ln tan 4 + ϕ )) t+1 (10) 2 ( ( π D t = Ln tan 4 + ϕ )) t (11) 2 R m = R e (1 2e +3esin 2 ϕ t+1 ) (12) where ϕ t and ϕ t+1 denote the latitudes of epochs t and t +1, respectively, λ t and λ t+1 denote the longitudes of epochs t and t + 1, respectively, R m is the semi-major axis of the earth, e is the eccentricity of the earth, S represents the step length between epoch t and t + 1, and A is the azimuth. Generally, the positioning error of PDR is based on two major types of errors: the along-track error mainly due to the step length assessment error and the cross-track error mainly due to the azimuth measurement error [9]. 5. Experiments and discussions The following section describes preliminary experimental results demonstrating the accuracy of position estimation using the wearable sensors by the proposed approach. Both the GPS coordinates and the wearable sensors data were collected and stored in a laptop. The data were post-processed, and the routes were calculated. The maps used in the experiments were from Google Earth, and they have been calibrated based on datum points. 5.1. Experiment 1 Latitude-N (degree) 31.203 31.2025 31.202 31.2015 31.201 31.2005 31.2 31.1995 B A GPS Starting point PDR True route PDR with fixed step length 121.591 121.592 121.593 121.594 121.595 121.596 Longitude-E (degree) Figure 11. The first experimental result. D C The first experiment was carried out on a road shaded by trees around an open rectangular area which was reserved as construction land. The aim of the experiment was to validate the feasibility of position estimation over a long distance and compare the performance of the proposed algorithm with that of the PDR algorithm using the fixed step length approach proposed in [1, 9]. In this experiment, the test subject covered a distance of 1440 m over a time period of 21 min and 5 s. During the experiment, 2175 steps were detected. The result is shown in figure 11; the green curve is the GPS route, the blue curve is the PDR route using the proposed algorithm, and the yellow curve is the PDR route with a fixed step length of 0.7 m which was also used in chapter 5 of [9]. The arrow indicates the direction of walking. The result revealed that the PDR route using the proposed algorithm fitted well with the true route. Its final offset was 31.5 m. The positioning error of the PDR route with a fixed step length was much larger; its final offset was 37.2 m. Since the step length of a pedestrian is a time-varying process, using a fixed step length will lead to along-track error accumulation which depends on how long the assumed fixed step length is between the markers A and B (highlighted in figure 11), the deviation of GPS positioning was very significant with an average error of 15.8 m, which may be due to the attenuation of the GPS signals by the foliage or the poor geometric distribution of satellites, while the average Figure 12. The azimuth measurements from the GPS receiver and built-in compass. error of the proposed approach was 4.8 m. It suggests that the accuracy of PDR using the proposed algorithm is higher than that of GPS over short distances; however, the accumulating error of PDR became quite large after a distance of 1110.3 m, with an average error of 22.1 m compared to that of GPS which was 4.9 m between C and D. Besides the accumulation of assessment errors of step length, anomalous outputs of the built-in electronic magnetic compass contributed to the across-track error accumulation of PDR. The azimuth measurements of the GPS receiver and builtin compass are shown in figure 12. The azimuth acquired from GPS deviated from that of the electronic compass between markers C and D. As shown in figure 12, the positioning accuracy of GPS is much higher than that of PDR around C and D, so it can be estimated that obvious drift of the azimuth measurements of the compass occurred resulting in the positioning error increasing rapidly. This deviation 8

Latitude-N (degree) 31.21 31.2098 31.2096 31.2094 31.2092 31.209 31.2088 31.2086 F GPS Starting Point PDR True route PDR without activity classification 121.5928 121.593 121.5932 121.5934 121.5936 121.5938 121.594 121.5942 121.5944 Longitude-E (degree) Figure 13. The second experimental result. in the azimuth may be the magnetic interference caused by an underground electric cable or the body offset of the subject. 5.2. Experiment 2 The second experiment presents a walk on a crowded street around a building which highlights the necessity of activity classification in the proposed PDR algorithm. The ground truth path was a 5 min and 49 s/434.21 m walk as shown in figure 13 which included a total of five right angled turns. During the walk, 636 steps were detected. The average errors in the GPS, PDR using the proposed algorithm and PDR without activity classification were 11.08 m, 6.17 m and 10.6 m, respectively. The poor GPS performance may result from the multipath effect owing to the GPS signal reflection and refraction caused by the building complex. In the distance between E and F highlighted in figure 13, the results of the two PDR approaches both with and without activity classification differ more greatly; the average positioning error of the former was 10.4 m and the latter was 18.5 m. The PNN enabled our step counting algorithm to reject erroneous spikes in the accelerometer s output as the wearer attempted to avoid other pedestrians. The final positioning error of the PDR without activity classification is 23.9 m, while that of the PDR using the proposed algorithm is 14.7 m. This experiment indicates that activity classification is indispensable in the PDR algorithm when the subject walks in obstacle-filled environments. 6. Conclusions and future work An approach using wearable sensors for pedestrian navigation application is presented in this paper. The probabilistic neural network functions as an activity classifier to select useful acceleration samples for the step detection and step length assessment. Furthermore, raw acceleration and azimuth data E from the wearable sensors are preprocessed by a wavelet transform and a moving average filter, respectively. The step length is estimated based on the step frequency. Two experiments reveal the reliability of the methods. As researched in previous studies, multiple accelerometers attached on different parts of the body can improve the performance of the activity classification, so this will be a focus of future work. Our experiments have also revealed the susceptibility of electronic magnetic compasses to environmental interference, particularly in an urban environment, so a gyroscope solution will be adopted in future trials. Moreover, the integration of sensors with GPS which is not described in this paper will certainly improve the performance of PDR, so the longer term aim is to couple GPS with wearable sensor using data fusion techniques, such as the Kalman filter. Acknowledgments This project would not have been possible without the help of Hao Wang, Kezhi Zhang, Guoping Xu, Shuwen Dang, Jian Li and Xifeng Yao, who inspired us with great ideas and helped us in experiments. We would also like to thank Lei Ji and Simon O Callaghan (Australian Centre for Field Robotics, The University of Sydney) for polishing the language. We are also very grateful to the anonymous paper reviewers for their helpful comments, suggestions and references. References [1] Randell C, Djiallis C and Muller H 2003 Personal position measurement using dead reckoning Proc. 7th IEEE Int. Symp. on Wearable Computers pp 166 73 [2] Jirawimut R, Ptasinski P, Garaj V, Cecelja F and Balachandran W 2003 A method for dead reckoning parameter correction in pedestrian navigation system IEEE Trans. Instrum. Meas. 52 209 15 [3] Levi R W and Judd T 1999 Dead reckoning navigational system using accelerometer to measure foot impacts US Patent 5583776 [4] Lee S W and Mase K 2001 Recognition of walking behaviors for pedestrian navigation Proc. 2001 IEEE Int. Conf. Control Applications (CCA 01) pp 1152 5 [5] Ladetto Q, Gabaglio V and Seeters J V 2004 Pedestrian navigation method and apparatus operative in a dead reckoning US Patent US6826477, Ecole Polytechnique Federale de Lausanne (EPFL) [6] Ladetto Q and Merminod B 2002 In step with INS: Navigation for the blind, tracking emergency crews GPS World Oct. 30 8 [7] Ladetto Q and Verhaert K 2007 Pedestrian navigation apparatus and method USPatent US20070260418A1, Vectronix AG [8] Ladetto Q 2000 On foot navigation: continuous step calibration using both complementary recursive prediction and adaptive Kalman filtering ION GPS 2000 1735 40 [9] Mezentsev O 2005 Sensor aiding of HSGPS pedestrian navigation PhD Thesis Department of Geomatics Engineering, University of Calgary [10] Lee S W and Mase K 2001 A personal indoor navigation system using wearable sensors Proc. ISMR (2nd Int. Symp. on Mixed Reality) pp 147 8 9

[11] Bao L and Intille S S 2004 Activity recognition from user-annotated acceleration data Proc. 2nd Int. Conf. Pervasive Computing pp 1 17 [12] Van Laerhoven K and Cakmakci O 2000 What shall we teach our pants? 4th Int. Symp. Wearable Computers pp 77 83 [13] Carter T 2006 An introduction to information theory and entropy Complex Systems Summer School pp 22 34 [14] Specht D F 1988 Probabilistic neural networks for classification, mapping, or associative memory IEEE Int. Conf. Neural Networks vol 1 pp 525 32 [15] Wang H, Zhang B and Tian W F 2007 Head tracking by means of probabilistic neural networks Meas. Sci. Technol. 18 1999 2004 [16] Daubechies I 1996 Where do wavelets come from? A personal point of view Proc. IEEE 84 510 3 [17] Bauby C E and Kuo A D 2000 Active control of lateral balance in human walking J. Biomech. 33 1433 40 [18] Ma Y S 2007 Design of a navigation computer based on ARM for vehicle navigation system MD Thesis Navigational and Control Lab, Shanghai Jiao Tong University 10