Segmentation :
Segmentation is the process of dividing the useful information from raw data since it'''s really hard to get helpful information from continious stream data. Thus, signal will be splited to meaningful pieces to enhance signal behaviour. For this purpose, different segmentation techniques can be grouped into one of the following three categories which are sliding windows, top-down and bottom-up on raw data. Sliding Windows is the most applied approach within segmentation techniques. It is uncomplicated and slight. In this technique streaming data is seperated into equal and diversed sized parts. The main difficulty with this technique is deciding the convenient size for every block. Top-down is an another segmentation technique. This method recursively divides the time arrangement until some stopping criteria is met. Bottom-Up technique, on the other hand, beginning from the finest conceivable estimate merges the segments until some stopping criteria is met.
Classification:
Classification is the final and the most significant step of the activity recognition process. The classication is the issue of recognizing to which of an arrangement of classifications another perception has a place, on the premise of a preparation set of information containing perceptions whose class participation is known Even though there are many, the most well-known classification methods are K-Nearest Neighbors (KNN), Decision Tree (DT), Support Vector Machines (SVM), Na''ve Bayes (NB), Hidden Markov Model (HMM) and Gaussian Mixture Models (GMM).
Related Work
In this chapter, we review previously proposed methods which use accelerometer or/and gyroscope information to detect human itself, human activities, stress, fall etc. Since it is getting more popular, many researches can be found about getting accelerometer or/and gyroscope information using smartphone sensors. Common usage of smartphones with much better sensors and technology draw researchers''' attentions to this area of study. While creating great input for our study, all those studies enlighten our research as well.
3.1 Previous Work
Randell and Muller added activity sensor to their GPS based Tourist Guide with a single X ''' Y accelerometer using the crossbow ADXL202 Accelerometer Evaluation Board in 2000. The data was collected from 10 people with activities; walking, running, sitting, walking upstairs, downstairs, and standing at a relatively low frequency 5Hz. User'''s action was inferred by using a grouping algorithm and a neural system. Initial results were 85 ''' 90 % with a high precision [4].
Mantyjarvi et al. used multiple acceleration sensors with best classification results 83 ''' 90 % by training three multilayer perception neural networks using back propagation (MLP) in 2001. For this purpose, they tested the use of PCA and ICA in feature generation process with wavelet transform [5].
In 2004, Bao and Intille gather acceleration data for their work from 20 people (13 males and 7 females) who run in age from 17 to 48 (mean 21.8, sd 6.59) with accuracy rate of 84%. It was the first study, to examine performance of recognition algorithms using many different accelerometers without wire. ADXL210E accelerometers were for collecting data from analog devices. Components were figured on 512 example windows of acceleration information with 256 tests covering between back to back windows at a sampling frequency of 76.25 Hz, each window represents 6.7 seconds. C4.5 decision tree and na''ve Bayes classi'''ers in The Weka Machine Learning Algorithms Toolkit were used for this study [6].
In 2005, Ravi et al. reported on their efforts to recognize user activity from accelerometer data. For this purpose they collected data with a sensor. To collect data, they selected 8 activities which are standing, walking, running, sit-ups, vacuuming, brushing teeth climbing upstairs, and downstairs. Sample frequency was 50Hz. Data were collected by the sensor which was triaxial accelerometer CDXL04M3. Subject'''s data transmitted to an HP iPAQ wirelessly over Bluetooth from the sensor. The data were converted to ASCII format using a Python script. Features were extracted from the raw accelerometer data using a window size of 256. Each window overlapped with the previous one with half the size of a window. The extracted features were mean, standard deviation, energy and correlation. They used WEKA toolkit for classification. Classifiers are Decision Tables, Decision Trees (C4.5), K-nearest neighbors, SVM, Naive Bayes, Stacking with MDTs, Stacking with ODTs, Plurality Voting, Bagging, and Boosting [7].
VTT Electronics workers proposed using identify people from accelerometer data in 2005. They used Analog Devices ADXL202JQ as a sensor for collecting data. The position of the sensor was behind the people'''s back. Their test subjects were 36 people (19 males and 17 females). The signals frequency was 256Hz. As a result, they got FRR=%5.4 and FAR= %6.4 [8].
Fitzgerald Nowlan published a paper (in 2009) about how human identification via gait recognition. For this purpose, he uses single sensor which is composed of an accelerometer and gyroscope and collected gait characteristic data. K-nearest neighbor, naive bayes and quadratic discriminant analysis selected for classification. In this work, the test subjects were able to be identified with 95% accuracy [9].
The idea of the implementation of a real-time classification system for some basic human movements using a conventional mobile phone equipped with a accelerometer and without server processing data was presented by Brezmes et al. in 2009. Nokia N95 cell phone was used for the prototype and Python API was used to obtain the accelerometer'''s data. This study based on activities such as walking, climbing-down stairs, climbing-upstairs, sitting down, standing up, falling. Albeit most studies on subject activities' acknowledgment utilize a few accelerometers situated at specific body destinations and with particular introductions, in this study the cell phone is hold by any user with no predefined introduction [10].
Spranger and Zazula make gait identification using cumulants of accelerometer with cell phone in 2009. Test set includes six males whose average ages were 30.2 years and average height 179. Experiment was performed on a 50 m long corridor with the surface made of stone plates. Nokia N95 was selected as a smartphone which were placed on the person'''s hip. Sampling data frequency is 37Hz. Feature extraction from detected gait cycles using cumulants. 1641 feature vectors generated by all cumulant coefficients from zero-lag cumulant to cumulant with lag 10. The classification was provided by support vector machines using WEKA. The average success rate was 93.1% [11].
In 2010, Kwapisz et al. published a paper about how a smart phone can be used to perform person identification and authentication. They used WISDM which is a cell phone platform based on android to collect data. For their work, they acquisitioned data from 36 users who performed four activities such as walking, jogging, climb up and down stairs. Users carried smart phones at their front pants''' pocket. Different types of Android phones were used such as Nexus One, HTC Hero and Motorola Backflip for this experiment. Even though example duration were tested for 10 seconds and 20 seconds but only 10 seconds data was selected since it'''s more reliable. Sampling data frequency level was 20Hz. Features were generated from 600 raw accelerometer data. 43 feature vectors were generated from variations of six based feature vectors that are average, standard deviation, average absolute difference, Average Resultant Acceleration, Time between Peaks and Binned Distribution. Two classification techniques were used on WEKA which is J48 and neural networks for data mining
Rasekh et al. published their paper in 2011 about human activity like walking, limping, jogging, going upstairs, and downstairs recognition system based on a 3 dimensional smartphone accelerometer. As a smartphone they used HTC Evo. Activities were classified, tested and trained using 4 different passive learning methods as quadratic classifier, k-nearest neighbor algorithm, support vector machine and artificial neural networks. Maximum sampling frequency of accelerometer was 50Hz and +/- 3g sensitivity. 31 features in both time and frequency domain were generated by the accelerometer with the system gathered time series signals. These time domains are variance, mean, median, 25% percentile, 75% percentile, Correlation between Each Axis, Average Resultant Acceleration (1 resultant failure) and frequency domain are energy, entropy, centroid frequency, peak frequency [13].
In 2011, Weiss and Lockhart published paper which is '''Identifiying User Traits by Mining Smart Phone Accelerometer Data'''. They used smartphones to predict user trait with accelerometer data. Their traits consisted of sex, height and weight. Data was collected by their WISDM Sensor application. Example duration was 10 second and frequency was 50 ms. Data set involved 66 subjects for gender prediction, 61 subjects for height prediction, 63 subjets for weight prediction. Feature vector was extracted average, standard deviation, average absolute difference, average resultant acceleration, time between peaks and binned distribution. They used WEKA and classification methods were Instance Based (IB3), Neural Network (NN) and Decision Tree (J48). The accuracy of sex prediction was 71.2% at IB3 classification method, accuracy of weight prediction was 78.9% at IB3 classification method, accuracy of height prediction was 85.7% at NN classification method [14].
In 2014, Ferrer and Ruiz compared different algorithms for the acknowledgment of transportation modes in view of elements removed from the accelerometer information. Android application called PEATON that is able to gather GPS readings each 10-12 seconds and accelerometer data at 1Hz. They used following smartphone models for their tests; Sony Xperia U, SonyXperia ArcS, Samsung Galaxy S, Samsung Galaxy S II and Google Nexus S. Information was gathered by 7 people over a time of three months: 4 male and 3 female, between the ages of 25 and 38. Members were told to completely select the travel mode picked while beginning an outing (walk, bicycle, motorcycle, car, bus, electric tramway, metro, train, or wait if participant is transferring between transport modes). The length of the sliding window is set as 30 seconds without covering between continuous windows. Five models; (1) k-Nearest Neighbors (KNN), (2) Decision Trees (DT), (3) Discriminant Analysis (DA), (4) Multilayer Perceptron Neural Network (NN), (5) Recurrent Neural Network (RNN) were used in Matlab for comparing the precision accuracy [15].
Celenli et al. used smart phones to detect activities while performing a certain action in 2014. The activities consisted of 7 basic and 1 complex actions; walking, running, jumping, standing, ascending stairs, descending stairs, standing up and sitting down as one action, getting in and out of a car. Each activity was 30 seconds. Subject database include 102 persons (35 of them were females and 67 of them were males) for basic actions. Subject'''s age average was 30. 30 subjects performed for complex action. IOS application was developed which includes accelerometer and gyroscope sensors. iPhone 5 and iPhone 5s smart phones were used in this approach so as to data acquisition. Frequency was chosen 100 Hz. Phone locations or the place of the action weren'''t specified. C++ code was developed for feature generation from collected data. Feature vector was extracted min, max, mean, Standard Deviation, Root Mean Square, Zero Crossings, Binned-average. They used WEKA toolkit classifiers which include regression, Bagging, Multi-layer Perceptron, K-Star ,Bayesian Network, Logistic Model Tree for classification. As a result, K-star led to recognition rates exceeding 98% [16].
In 2014 Aguiar et al. developed Android application which is ADLS for fall detection using accelerometer data. When a fall was detected, sound the alarm of application. Their test subject set includes 36 people, 28 of them are young people and 8 of them are older people. 24 males and 4 females of young people with average age of 25, average height of 175, average weight of 71. 4 males and 4 females of older young people with average age of 66, average height of 175, average weight of 72. They tested three different classifiers of offline machine learning tool: Decision Trees, K-Nearest-Neighbors and Naive Bayes. They obtained feature vector from mean, median, maximum, minimum, root mean square, standard deviation, median deviation, interquartile range, energy, entropy, skewness and kurtosis. The success of their fall detection algorithm was 97.5% [16].
Jain and Kanhangad proposed, in 2015, a method using biometrics for user authentication which achieves the lowest EER of 0.31%. There were 104 participants. The dataset comprises 9 users between the age 31and 36 years, 40 users between 26 and 30 years and 55 users between 19 and 25 years. IntelliJ IDEA android application was developed and ran it on Samsung Galaxy S-II GT-I9100. Samsung Galaxy Note-II N7100 was used as well to acquire dataset that contains data from 30 subjects [17].
Osmani et al. proposed in 2015 to use smartphones for detecting behavior changes from accelerometer data. Their test has been done on 6 persons for 10 months at psychiatric hospital in Hall in Tirol, Austria. The authors reported that they could gain success average precision was 81%, and recall was 82% by using Naive Bayes, k-nearest neighbor, j48 search tree and a conjunctive rule learner. They evaluated the test with the Gaussian distribution method and increased the success rate average precision to 96%, and recall to 94% [18].
In 2015, Garcia-Ceja et al. achieved a maximum overall accuracy of 71% for user-specific models and an accuracy of 60% for the use of similar-users models with 30 subjects (18 [60%] males and 12 [40%] females) for automatic stress detection in working environments from smartphones''' (using the built-in sensors of Samsung Galaxy SIII Mini smartphones) accelerometer data. In order to extend the battery life, they set the accelerometer sampling rate at 5 Hz. They used 4 classifiers: Naive Bayes; Decision Tree; Ordinal Naive Bayes for their experiments. Accelerometer information was used to portray subjects' conduct by extricating time domain and frequency domain [20].
Another paper that was published in 2015 by San-Segundo et al. aimed to get feature extraction from smartphone inertial signals for human activity segmentation. Those segmentations were six different physical activities; walking, walking'''upstairs, walking-downstairs, sitting, standing and lying. Input data was ensured 30 volunteers. This dataset has been randomly divided into sixsubsets for performing a six-fold cross validation procedure. The dataset which was called UCI Human Activity Recognition Using Smartphones dataset shared publicly. They stated that Activity Segmentation Error Rate as lower than 0.5%. It was recommended that defining new frequency warping strategies and focusing on evaluating these feature extraction proposals in a different application: gait recognitioninstead of activity recognition as future work.
Biometric Authentication Technique Using Smartphone Sensor was published in 2016 by Laghari, Rehman and Memon. The paper presents authenticating biometrically with the help of smartphone'''s motion sensor. Signal matching concept was used for identification. They concluded their experiment result as 6.87% FRR and 1.46% FAR. As they mentioned the method they used can be improved and more accurate results can be gathered from frequency analysis of the signature signal for authentication process.
To compare ours to previous works can help describing and explaining the differences beetween studies for clear and better understanding and also it may point new research ideas. To achive this Table I was created. At this table, columns are corresponded as follows; reference to the work, sensor type, using device, proposed method, the number of subjects data was collected from, sampling rate, overall best success among different classifiers''' results, feature vectors and classifiers, respectively.
There are 4 main contributions of our work. Firstly, we collected data from a large set of subjects with varying ages and gender. Secondly,data were acquisitioned from subjects at different environment.Thirdly, application were developed at two different platform which are IOS and Android for getting data. Lastly, more than one methods were offered.