High Five: Improving Gesture Recognition by Embracing Uncertainty

10/25/2017 ∙ by Diman Zad Tootaghaj, et al. ∙ The University of Texas at Austin Microsoft cornell university 0

Sensors on mobile devices---accelerometers, gyroscopes, pressure meters, and GPS---invite new applications in gesture recognition, gaming, and fitness tracking. However, programming them remains challenging because human gestures captured by sensors are noisy. This paper illustrates that noisy gestures degrade training and classification accuracy for gesture recognition in state-of-the-art deterministic Hidden Markov Models (HMM). We introduce a new statistical quantization approach that mitigates these problems by (1) during training, producing gesture-specific codebooks, HMMs, and error models for gesture sequences; and (2) during classification, exploiting the error model to explore multiple feasible HMM state sequences. We implement classification in Uncertain<t>, a probabilistic programming system that encapsulates HMMs and error models and then automates sampling and inference in the runtime. Uncertain<T> developers directly express a choice of application-specific trade-off between recall and precision at gesture recognition time, rather than at training time. We demonstrate benefits in configurability, precision, recall, and recognition on two data sets with 25 gestures from 28 people and 4200 total gestures. Incorporating gesture error more accurately in modeling improves the average recognition rate of 20 gestures from 34% in prior work to 62%. Incorporating the error model during classification further improves the average gesture recognition rate to 71%. As far as we are aware, no prior work shows how to generate an HMM error model during training and use it to improve classification rates.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

I Introduction

Modern mobile devices host a diverse and expanding array of sensors: accelerometers, gyroscopes, pressure meters, thermometers, ambient light sensors, and more. These sensors invite new experiences in fitness, health, translating sign language, games, and accessibility for people with disabilities [1, 2, 3]. Despite all these new input methods, user input on smartphones is still mostly limited to touching the screen and keypad, a 2-D detection problem. This paper identifies and addresses algorithmic and practical impediments to deploying 3-D gesture recognition on smartphones. We extend the commonly used Hidden Markov Model (HMM) approach [4, 5, 6, 7]. Determining a 3-D path through space is harder than 2-D gesture recognition [8] because human gestures as captured by sensors are uncertain and noisy—much noisier than the sensors themselves. Humans hold the device at different angles, get tired, and change their gestures’ pattern. Prior state-of-the-art gesture-recognition algorithms using HMMs [6, 9, 10] are limited because (1) they assume all gesture error is uniform and project all observations to one spherical codebook for HMM training; and (2) classification generates one observation sequence and produces only one deterministic gesture, rather than reasoning explicitly about the uncertainty introduced by gesture error.

We measure gesture noise in accelerometer data and find it is a gesture-specific Gaussian mixture model: the error distributions along the x, y, and z axes all vary. In contrast, when the phone is still, accelerometer error is extremely small, Gaussian, and uniform in all dimensions. Gesture-specific error matches our intuition about humans. Making an “M” is harder and more subject to error than making an “O” because users make three changes in direction versus a smooth movement. Even when making the same gesture, humans hold and move devices with different orientations and rotations. Since gesture observation is a sequence of error readings, differences in gesture sizes and speed can compound gesture error.

State-of-the-art HMM systems [6, 9, 11] assume errors are small, uniform, and not gesture specific. They compute one radius for all gestures and all x, y, and z accelerometer data. They map all gestures to a single spherical set of codewords centered at that they use to train the HMM. Classification compounds this problem because HMMs use deterministic quantization

. Even though several nearby states may be very likely, traditional HMM classifiers only explore one.

To solve these problems, we present a holistic statistical quantization approach that (a) computes and reasons about noise gesture training data; (b) produces per-gesture HMMs and their error models; (c) modifies classification to use the error model to choose the most likely gesture; and (d) uses the probabilistic programming system [12, 13]

to simplify the implementation and expose the classifier’s trade-off between precision and recall.

During training, we measure error in accelerometer data sequences across gestures and use the mean and variance to improve HMM modeling and classification. In training, we fit per-gesture data to codewords on an ellipse and generate gesture specific HMM codebooks. We show that ellipse-based codebooks improve accuracy over prior sphere-based approaches 

[6, 9]. We target personal mobile devices where users both specify and train gestures. With per-gesture HMM models, users train one gesture at a time. Instead of performing classification by deterministically mapping the 3-D acceleration data to the closest codeword, we sample from the error model produced during training to explore a range of potential gestures.

We implement classification as a library in the

programming language. The library provides trained HMM models and their error models. A gesture is an Uncertain type. Values of Uncertain types represent probability distributions by returning samples of the base type from the error distribution. The runtime lazily performs statistical tests to evaluate computations on these values. When the application queries an Uncertain value, such as with an

if

statement on the gesture, the runtime performs the specified statistical hypothesis test by sampling values from the HMM computation.

We evaluate statistical quantization on two data sets: (1) five gestures trained by 20 people (10 women and 10 men) on a Windows Phone that we collect, and (2) 20 gestures trained by 8 people from Costante et al. [14]. Compared to traditional deterministic spherical quantizers [6], statistical quantization substantially improves recall, precision, and recognition rate on both data sets. Improvements result from better modeling and using error in classification. Deterministic elliptical quantization improves average gesture recognition rates on 20 gestures to 62%, compared to the 34% for traditional deterministic spherical quantization. Statistical elliptical quantization further improves gesture recognition rates to 71%.

We illustrate the power of our framework to trade off precision and recall because it exploits the error model during classification. Prior work chooses one tradeoff during training. Different configurations significantly improve both precision and recall. This capability makes statistical quantization suitable both for applications where false positives are undesirable or even dangerous, and for other applications that prioritize making a decision over perfect recognition.

Our most significant contribution is showing how to derive and use gesture error models to improve HMM classification accuracy and configurability. Our

approach is a case study in how a programming language abstraction for error inspires improvements machine-learning systems. HMM inference algorithms, such as Baum–Welch, exemplify software that ignores valuable statistical information because it can be difficult to track and use. They infer a sequence of hidden states assuming

perfect sensor observations of gestures. Our approach invites further work on enhancing inference in other machine learning domains, such as speech recognition and computational biology, that operate on noisy data. We plan to make the source code available upon publication. The compiler and runtime are already open source [13].

Ii Overview of Existing Approaches

Fig. 1: Design space for gesture recognition algorithms.

Recognizing human gestures is key to more natural human-computer interaction [15, 6, 16, 17]. Sensor choices include data gloves [1], cameras [18], touch detection for 2-D painting gestures [8], and our focus, 3-D motion tracking accelerometer and gyroscope sensors [19]. Figure 1 presents the design space for common gesture recognition approaches. Non-Probabilistic approaches include dynamic time warping [16, 17, 11, 20, 21]

and neural networks

[22]. A common probabilistic approach is Hidden Markov Models (HMMs) [6, 9, 23, 10] that use non-linear algorithms to find similar time-varying sequences.
Hidden Markov Models: HMMs for gesture recognition give the best recognition rates for both user-dependent and user-independent gestures [24]

. Our HMM implementation for gesture recognition differs from the prior literature as follows. First, instead of using a deterministic codebook for discretization, we use the statistical information about each gesture during HMM training and generate a different codebook for each gesture. Second, we exploit a probabilistic programming framework in our implementation and use uncertain data types to make more accurate estimations of the probability of each gesture. Third, unlike prior work that deterministically maps raw data to one static codebook, we use a stochastic mapping of raw scattered data based on the gesture’s error model and the distance from the data to each gesture’s trained codebook.


Kmeans quantization: Since continuous HMMs for gesture recognition is impractical due to high complexity of tracking huge observation states, a variety of quantization techniques transform sensor data into discrete values. The most-common is kmeans clustering [6, 25]. Although kmeans

works well for large isotropic data sets, it is very sensitive to outliers and therefore noisy sensor data degrades its effectiveness. For example, a single noisy outlier results in a singleton cluster. Furthermore, because humans must train gesture recognizers, the training gesture data sets are necessarily small, whereas

kmeans is best suited for large data sets.
Dynamic time warping

: Dynamic time warping applies dynamic programming to match time-varying sequences where gesture samples are represented as feature vectors

[16, 17, 24]. The algorithm constructs a distance matrix between each gesture template and a gesture sample . The algorithm next calculates a matching cost between each gesture template and the sample gesture. The sample gesture is classified as the gesture template with the minimum matching cost. This approach is easy to implement, but its accuracy for user-independent gestures is low [11, 26]. Furthermore, it is deterministic and does not capture the stochastic and noisy behavior of accelerometer and gyroscope’s data.
Neural networks: Neural networks classify large data sets effectively [22]. During training, the algorithm adjusts the weight values and biases to improve classification. While neural networks work well for large data sets, their applicability is limited for small amounts of training data [5]. Asking end users to perform hundreds of gestures to train a neural network model is impractical.

Fig. 2: HMM example with state and output probabilities.

Iii Hidden Markov Model Background

HMMs are used extensively in pattern recognition algorithms for speech, hand gestures, and computational biology. HMMs are Bayesian Networks with the following two properties: 1) the state of the system at time t (

) produces the observed process (), which is a random or deterministic function of the states, is hidden from the observation process, and 2) the current state of the system given the previous state is independent of all prior states for [4]. The goal is to find the most likely sequence of hidden states. Generally speaking, an HMM is a time sequence of an observation sequence , derived from a quantized codebook , that is . In addition, hidden states are derived from the states in the system , that is . The state transition matrix models the probability of transitioning from state to . Furthermore, models the probability that the hidden state generates the observed output .
Figure 2 shows an example of an HMM model with two hidden states, three observed states, and the corresponding state transition and output probabilities. This HMM is ergodic because each hidden state can be reached from every other hidden states in one transition. A left-to-right HMM model does not have any backward transitions from the current state. We consider both ergodic and left-to-right HMM models.

Iv Limitations of Existing Geture Recognition Approaches

In theory, all machine learning algorithms tolerate noise in their training data. Common approaches include using lots of training data, adding features, and building better models, e.g., adding more interior nodes to an HMM. In practice, we show that understanding and measuring error inspires improvements in modeling and classification.

Fig. 3: Location errors accumulate from various small angle errors.

Iv-a 3-D Path Tracking is a Hard Problem [27]

A gesture is a meaningful 3-D movement of a mobile device. When the phone moves, sensors gather an observation consisting of a sequence of 3-D accelerometer data. We use accelerometer observations to train 3-D gesture recognition models, but our approach should apply to other sensors.

Since users hold devices at different angles and move at different velocities even when making the same gesture, the 3-D accelerometer data includes a gravity component on all axes, which varies per user and gesture. Figure 6 shows how the phone angle generates different acceleration data projected on the X, Y, and Z axes due to gravity. One approach we tried was to eliminate gravity variation by using gyroscope data and tracking a 3-D path. This approach does not work because it actually amplifies error. We show this derivation to motivate the difficulty of path tracking and how gesture errors make it worse.

A 3-D path tracking approach models total acceleration and projects it to a position as follows. Given

(1)

where is the measured data from the accelerometer; is the actual acceleration applied to the phone’s frame by the user; is the rotation matrix between the actual force applied to the phone and the frame of the sensor; and is a unique vector along z direction [28, 29]. Rotating the sensor frame acceleration to the actual force frame gives the inertial acceleration:

(2)

Integration of the inertial acceleration produces a velocity and then integrating acceleration twice produces a phone position.

(3)
(4)

A rotation matrix is obtained by multiplying each of the yaw, roll, and pitch rotation matrices. Adding gyroscope data, and assuming the phone is still at (which means we know the initial angel with respect to gravity), the accumulated rotational velocity determines the 3-D angles with respect to gravity at any time [28, 29, 30]. Projecting the accelerator data in this manner may seem appealing, but it is impractical for two reasons. (1) It results in dimensionless gestures, which means the classifier cannot differentiate a vertical circle from a horizontal circle. Users would find this confusing. (2) It amplifies noise making machine learning harder. Figure 3 shows the accumulated error over time for different values of angle error. Even small errors result in huge drift in tracking the location, making gesture tracking almost impossible. Consequently, we need to use a different approach to handling gesture errors.

Fig. 4: Error distribution of accelerometer data for “O”.
Fig. 5: Error distribution of accelerometer data for “N”.
Fig. 4: Error distribution of accelerometer data for “O”.
Fig. 6: The phone angle changes accelerometer readings.

Iv-B Noise is Gesture Specific

We collect the error distribution, mean, and variance of the x, y, and z accelerometer readings for each gesture at each position in sequence. Figures 5 and 5 plot the resulting distributions for two examples from the “O” and “N” gestures. The error distributions tend to be Gaussian mixture models since the accelerometer measure x, y, and z coordinates. Because the error is large and differs per gesture, it suggests different models for each gesture should be more accurate. The error distributions are not uniform. Mapping the data to codewords can exploit this information. Prior approaches fall down on both fronts: they do not learn separate models for each gesture or accomodate gesture noise in their codeword maps [4, 5, 6, 7]

Iv-C Noise in Classification

Noise affects how the system maps a sequence of continuous observations to discrete codewords. A deterministic quantization algorithm does not deal with this source of error. Figure 7 illustrates why deterministic quantization is insufficient. The black points are codewords and the white point is a sensed position in 2D space. The distances and

are similar, but the B codeword is slightly closer, so deterministic quantization would choose it. In reality, however, the sensed value is uncertain and so too are estimates of the codewords themselves. The gray discs show a confidence interval on the true position. For some values in the interval,

and thus the correct quantization is A.

Our statistical quantization approach explores multiple codewords for the white point by assuming that points close to the sensed value are likely to be the true position. In other words, the probability of a given point being correct is proportional to its distance from the sensed value. We therefore choose the codeword A with a probability proportional to and B with a probability proportional to .

Fig. 7: Error in sensor and codeword estimation (shaded circles) is mitigated by .

V High Five: Gesture Training and Classification

To accurately model and recognize gestures, we propose two techniques: deterministic elliptical quantization and statistical elliptical quantization.
Deterministic Elliptical Quantization During training, we gather the statistical data on errors (distribution of error, mean, and variance) for each position in gesture sequences, create codewords, and train HMMs for each gesture. We map all the observation sequences for each gesture to a unique codebook. We construct an elliptical contour for our codebook based on mean and variance of the observations. Figure 8 shows the spherical equal-spaced codebook generated for all gestures based on prior work [6] and our per-gesture ellipses for three example gestures. In the figure, the acceleration data is expressed in terms of . If we hold the phone along the Z axis, the scattered data has a bias of which shows the gravity component. If the user holds the phone upside-down the scattered data has a bias of and our statistical generated codewords embraces the gravity component in each case. Per-gesture ellipses better fit the data than a single sphere for all gestures. We use 18 equally spaced points on the elliptical contour to ease comparison with related work, which uses 18 points on a spherical contour [6]. 18 observation states strikes a balance between learning complexity and accuracy, both of which are a function of the number of states. This method is similar to multi-dimensional data scaling [31], but as we showed in the previous section, standard projection is a poor choice for this data. We construct elliptical models for each gesture as follows.

(5)

The values , and are the expected value of raw acceleration data for each gesture. We construct a different codebook for each gesture. This process maps the accelerometer data to one of the 18 data points as shown in Figure 8. The mapped data constructs the observed information in the Hidden Markov Model.

Our quantization approach differs from the prior work [6] in two ways. First, since we use the statistics of each gesture, there is no need to remove the gravity bias, because the center of mass for all gesture data of a specific gesture includes the gravity component. The second difference is that we chose a different contour for each gesture in our data set. As Figure 8 shows, the elliptical contour for a x-dir gesture is completely different from the contour for y-dir or N.In the spherical contour, most of the data points from the accelerometer map to a single codeword, eliminating a lot of information that is useful for classification. Our approach reduces the quantization error for different gestures since it is much less likely to map each gesture to another gesture’s codebook and generate the same sequence.

Fig. 8: Eliptical and spherical quantization for 3-D accelerometer data for “x-dir”, “y-dir”, and “N” gestures.

Gesture Training After measuring the noise and training the codebooks for each gesture, we build an HMM model for each gesture. The gesture recognition application takes as input the 3-D accelerometer sequence for each gesture and updates the HMM probabilities using the forward-backward algorithm [32, 6, 5, 15]. We use Baum–Welch filters to find the unknown parameters of each gesture’s HMM model (i.e., and ) [33]. Assuming is independent of time and assuming the probabilities of initial states is , the probability of a certain observation at time for state is given by

(6)

Baum–Welch filters use a set of Expectation Maximization (EM) steps. Assuming a random initial condition

for the HMM, Baum–Welch finds the local maximization state transition probabilities, output probabilities, and state probabilities. That is, HMM parameters which will maximize the observation probabilities as follows.

(7)

The algorithm iteratively updates , , and to produce a new HMM with a higher probability of generating the observed sequence. It repeats the update procedure until it finds a local maximum. In our deployment, we store one final HMM for each gesture as a binary file on the phone.

Statistical Gaussian Mixture Model (GMM) Quantization The key to our Statistical Elliptical Quantization

approach is representing each gesture reading as a random variable that incorporates its sensor noise. For classification, our statistical quantization approach uses the Gaussian distribution mixture models based on the error model we observe for each gesture during training. For example, Figures 

5 and 5 show that the probability distribution of distance of mapped to follows a Gaussian mixture model distribution with 3 peaks each representing the peak over one of the X, Y or Z coordinate. The probability of mapping a data point to each codeword for a bivariate Gaussian noise distribution is computed as follows:

(8)

A mixture of three Gaussian distribution models maps to individual Gaussian models as follows:

/* Quantization code for a mixture of three Gaussian distribution. */
var acc = ReadAccelerometer();
var d = /*@ $\|{(acc, codeword_i)}\|$ @*/ ;
if /*@ $ d < ( \mu_1 + \mu_2)/2$ @*/ :
/*@ $d = N( \mu_1, \sigma_1 )$ @*/ ;
else if /*@ $ ( \mu_1 + \mu_2)/2 < d < ( \mu_2 + \mu_3)/2 $ @*/ :
/*@ $ d = N( \mu_2, \sigma_2)$ @*/ ;
else:
/*@ $d = N( \mu_3, \sigma_3)$ @*/

This mapping produces a probability distribution over codewords for each reading. Sampling from this distribution creates multiple sequences of observation for the HMM, which then determines the most likely gesture from the entire distribution.

Statistical Random Quantization For comparison, we also implement a random quantizer to exploit error if training did not or was not able to produce a gesture-specific error model. This quantizer maps an observation randomly to codewords depending on their distance to the data point. For example, given four codewords, it randomly maps the gesture data proportional to the distance with respect to each codeword:

(9)

Gesture Classification The GMM quantization and Random quantization algorithms appear in Algorithms1 and 2, respectively. We implement these classifiers in the programming language (described below), exploiting its first-class support for probability distributions.

Algorithm 1 shows our novel statistical RMM quantization. Each step of the algorithm maps user data to a sequence of observation states from the generated codebook during training for each of gestures in . We treat the mapping independently for each data point in . (We also explored computing correlated mapping where mapping the current 3-D data to one of the quantization codewords depends on the previous 3-D mapping, which further improves accuracy, but for brevity omit it). At each step, we sample nearby codewords in and weigh them by their probability based on the GMM error model observed during training to create a sequence of observation states. We next classify the generated sequence to find , sample until the probabilities converge, and then pick the most likely sequence. When the algorithm completes, we have computed the most likely HMM path for each . We only consider with a probability above a threshold as potential gestures and thus may not return a gesture. For those above the threshold, we return the one with the highest probability as the most likely gesture. We explore values of 0.5 and and find that 0.5 works best.

Algorithm 2 uses the Random quantizer which implements a Bayesian classification scheme that returns the probability of each gesture given the observed sequence [34]. Given a set of observation sequence , it computes the probability of each gesture as follows.

(10)

The values and are produced by the Baum–Welch training model for each individual gesture.

Data: Raw accelerometer data, HMM models for , Inference threshold ()
Result: If or
1 Most probable gesture = ;
2 for   do
3       ;
4       while  or  do
5             X= Map Accelerometer data sequence to to ’s quantization codebook using random quantization;
6             Find from the Baum--Welch trained HMM model;
7             ;
8             Add to ’s distribution model;
9            
10      if  and  then
11             MaxProb = ;
12             Most probable gesture = ;
13            
14      
15return Most probable gesture;
Algorithm 1 High Five: GMM Quantization
Data: Raw accelerometer data, HMM models for
Result: The most probable gesture
1 MaxProb = 0 ;
2 Most probable gesture = ;
3 for   do
4       Map Accelerometer data to to the quantization codebook for each gesture using deterministic quantization;
5       Find from the Baum--Welch trained HMM model;
6       ;
7       if   then
8             MaxProb = ;
9             Most probable gesture = ;
10            
11      
12return Most probable gesture;
Algorithm 2 High Five: Random Quantization

Vi Discussion

Our HMM model tracks a single HMM path during classification but builds many possible input observations from a single trace of accelerometer data. In contrast, prior work shows it is possible to use an HMM which tracks the top paths during classification [35]

. It is interesting future work to explore any experimental differences in such a formulation. We use the raw accelerometer data as a feature given to our HMM training and classification. However, there exist prior works, especially in computer vision, that finds rotation-invariant or scale-invariant features

[36, 37]. We did not use rotation-invariant features because we want the capability to define more gestures, e.g., “N” in the x-y and in the z-y plane are distinct. However, more sophisticated features can further improve our classification accuracy.

Vii Implementation

To help developers create models for problems in big data, cryptography, and artificial intelligence, which benefit from probabilistic reasoning, researchers have recently proposed a variety of probabilistic programming languages

[12, 38, 39, 40, 41, 42]. We use the programming language to implement random quantization. We choose it, because is sufficiently expressive and automates inference, and thus significantly simplifies our implementation [12, 42, 43]. The remainder of this section gives background on , the programming model that inspired and supports our technique, and describes our implementation.

Vii-a The UncertainT Programming Model

is a generic type and an associated runtime in which developers (i) express how uncertainty flows through their computations and (ii) how to act on any resulting uncertain computations. To accomplish (i) a developer annotates some type as being uncertain and then defines what it means to sample from the distribution over through a simple set of APIs. Consumers of this type compute on the base type as usual or use LINQ primitives [44] to build derived computations. The runtime turns these derived computations into a distribution over those computations when the program queries it. Querying a distributions for its expected values and or executing a hypothesis test for conditionals, triggers a statistical test. Both of these queries free the runtime from exactly representing a distribution and let it rely on lazy evaluation and sampling to ultimately determine the result of any query.

Vii-B Statistical Quantization with UncertainT

/* Classification Code for Statistical Quantization. */
var acc = ReadAccelerometer();
Uncertain<int> gestures =
  // distribution over observations
  from obs in new StatisticalQuantizer(acc)
  // returns most likely gesture
  let gesture = Bayes.Classify(acc, obs)
  select gesture;
// T-test: more likely than not that
// this is the gesture labeled 0
if ((gestures == 0).Pr(0.5))
        Console.WriteLine("gesture=N");
Fig. 9: Statistical Quantization for a single gesture.

To implement statistical quantization, we express each gesture as a random variable over integer labels. Our implementation of StatisticalQuantizer(acc) (Figure 9) first reads from the accelerometer and passes this observation to the RandomQuantizer(acc) constructor which knows how to sample from observations by randomly mapping analog accelerometer data to discrete codewords to return a distribution over observations. The LINQ syntax let’s the developer call existing code designed to operate on type (i.e., Bayes.Classify which operates on concrete observations) and further lets her describe how to lift such computation to operate over distributions. In gesture recognition, the resulting type of gestures is then an Uncertain<int> or a distribution over gesture labels.

The runtime does not execute the lifted computations until the program queries a distribution’s expected value or uses it in a conditional test. For example, when the developer writes if ((gestures == 0).Pr(0.5)) the runtime executes a hypothesis test to evaluate whether there is enough evidence to statistically ascertain whether it is more likely than not that the random variable gestures is equal to the gesture with label 0. The runtime samples from the leaves of the program and propagates concrete values through any user-defined computation until enough evidence is ascertained to determine the outcome of a conditional. The

runtime implements many inference algorithms under the hood (rejection sampling, Markov Chain Monte Carlo, etc.). For this domain, we found no reason to prefer one over the other and so use rejection sampling for all experiments.

Fig. 10: Five gestures in Windows Phone (WP) data set.
Fig. 11: Twenty gestures in Smart-Watch (SW) data set [14].

Viii Methodology

This section describes the data sets, implementation details, and algorithms that we use to evaluate statistical quantization. We evaluate our algorithms and collect one data set on a smart-phone Nokia Lumia 920 (Windows Phone 8). We use the Windows SDK 8 API to read the 3-D accelerometer.

Data sets, training, and testing We collect our own data set on the Windows Phone and use the publicly available Smart Watch data set from Constante et al. [14], which together make a total of 4200 gesture samples of 25 distinct gestures performed by 28 people.

Windows Phone (WP) Data Set

We collect data from 10 men and 10 women performing 10 times each of the 5 gestures shown in Figure 10 on our Windows Phone platform, for a total of 1000 gesture samples.

Smart Watch (SW) Data Set

We also use a publicly available data set consisting of the 20 gestures shown in Figure 11 trained by eight people [14]. Each person performed each gesture 20 times for a total of 3200 samples.

Prior studies [6, 11] have smaller data sets and very distinct gesture patterns. In contrast, our data sets include gestures with very similar patterns. For example, W and N in the WP data set differ by about one stroke, and G9 and G11 differ by a 90 rotation, making them hard to differentiate.

These data sets represent a realistic amount of training for individuals, because users, even paid ones, are unlikely to perform the same gesture well 100s of times for training. Training must be short and recognition must be effective quickly to deliver a good user experience. To create sufficient training data, we train each classifier with data from all the users (20 for WP and 8 for SW). To assess the accuracy of the gesture recognition algorithms, we randomly split the data sets into 75% training data and 25% test data and repeat this procedure 10 times.

The High Five gesture recognition application We implement a Windows Phone gesture recognition application, called High Five. Users train the system online by first specifying a new gesture and then train the system by performing the gesture at least 10 times. Users can perform any gesture they want and then specify a corresponding action triggered by the gesture (e.g., call Mom on M, send an Email on E). The application has two modes: signaled, in which users open the gesture recognition application first before making a gesture, and dead start, which captures all device movements, and thus is more likely than signaled recognition to observe actions that are not gestures. We implement the system in Microsoft’s Visual Studio 2015 for C# and the Window’s Phone software development kit (SDK) sensor API. We use the libraries and runtime for our statistical quantizer by adding an HMM API to that returns samples from HMM distributions.

Gesture recognition algorithms We evaluate the following gesture recognition algorithms.

Deterministic Spherical Quantizer

Wijee is the prior state-of-the-art [6]. It uses a traditional left-to-right HMM with kmeans quantization and one spherical model for all gestures [6]. We follow their work by limiting transitions to four possible next codewords, such that is the only possible next state from where . (They find that left-to-right and ergodic HMMs produce the same results.) We extend their algorithm to train gesture-specific models using a unique codebook for each gesture. Since the scattered data is different for each gesture, using per-gesture codebooks offers substantial improvements over a single codebook for all gestures.

Deterministic Elliptical Quantizer

This algorithm uses a left-to-right HMM, elliptical quantization, and a unique codebook for each gesture.

Statistical GMM Quantizer

This algorithm uses a left-to-right HMM, statistical Gaussian mixture model (GMM) elliptical quantization based on observed error, and a unique codebook for each gesture. The runtime generates multiple observation sequences by mapping the data sequences to multiple codeword sequences for each gesture using a gaussian mixture model. With statistical quantization, the developer chooses a threshold that controls false positives and negatives, which we explore below.

Statistical Random Quantizer

This algorithm uses a left-to-right per-gesture elliptical HMM, statistical random quantization, and a unique codebook for each gesture.

(a) Deterministic spherical quantizer
(b) Deterministic elliptical quantizer
(c) Statistical GMM quantizer
Fig. 12: Precision and recall curves for gesture recognition algorithms.

Ix Evaluation

This section compares the precision, recall, and recognition rate of the gesture recognition algorithms. We show that statistical quantization is highly configurable and offers substantial improvements in accuracy, recall, and recognition over other algorithms. The other recognizers are all much less configurable and achieve lower maximum accuracy, recall, and/or recognition in their best configurations. These experiments illustrate that a key contribution of statistical quantization is that it has the power to offer both (1) highly accurate recognition in the signaled scenario, and (2) significant reductions in false positives in the dead-start scenario, thus matching a wide range of application needs.

We explore the sensitivity of gesture classification accuracy as a function of the number of gestures, using 2 to 20 SW gestures. For all the algorithms, accuracy improves with fewer gestures to differentiate. Statistical random quantization is however substantially more accurate than the others for all numbers of gestures. We further show that our approach is relatively insensitive to the number of users in the training data. Finally, we show how to easily incorporate personalization based on other factors, such as performing a subset of the gestures, and the result further improves accuracy.

Recognition rates for signaled gestures In this first experiment, users open the High Five application and then perform the gesture, signalling their intent. Figure 12 shows precision (dashed lines) and recall (solid lines) for each of the 5 gestures in distinct colors for the WP data set as a function of the conditional threshold. Precision is the probability that a gesture is correctly recognized and is an indication of false positives while recall shows the probability of recognizing a performed gesture and shows false negatives. The deterministic elliptical quantizer in Figure 12(b) uses the domain specific knowledge of each gesture during training and thus has higher precision and recall compared to deterministic spherical quantization in Figure 12(a).

Statistical GMM quantization in Figure 12(c) offers further improvements in precision and recall. Although the recall curve goes down as a function of the conditional threshold, when the conditional threshold is or lower, the recognition rate is higher than deterministic elliptical quantization. Statistical GMM delivers a similar threshold for precision. Statistical GMM quantization offers a distinct and smooth trade-off between precision and recall. Applications thus have a range of choices for the conditional threshold from which to choose that they can tailor to their requirements, or even let users configure. For instance, when the user is on a bus, she could specify higher precision to avoid false positives, since she does not want the phone to call her boss with an unusual movement of the bus. Prior work requires the training algorithm specify this tradeoff, instead of the end developers and users.

Figure 13

shows the recognition rate for each gesture in the WP data set for all the classifiers. The deterministic elliptical quantizer improves substantially over the deterministic spherical quantizer. Statistical GMM and random quantization deliver an additional boost in the recognition rate. Both GMM and random produce similar results within the standard deviation plotted in the last columns. On average, both statistical GMM and random quantization deliver a recognition rate of 85 and 88%, respectively, almost a factor of two improvement over deterministic spherical quantization.

Recognition rates for dead start and as a function of gestures This experiment explores the ability of the gesture recognition algorithm to differentiate between no gesture and a gesture since users do not signal they will perform a gesture. For instance, a gesture M must both wake up the phone and call your mom. In this scenario, controlling false positives when you carry your phone in your pocket or purse is more important than recall—you do not want to call your mom unintentionally.

Accuracy as a function of the number of gestures The more gestures the harder it is to differentiate them. To explore this sensitivity, we vary the number of gestures from 2 to 20 and compare the four approaches. Figure 14 shows the recognition rate for the deterministic spherical, deterministic elliptical, statistical random, and statistical GMM quantizers as a function of the number of gestures in the High Five application. All classifiers are more accurate with fewer gestures compared to more gestures. Increases in the number of gestures degrades the recognition rate of the deterministic spherical faster than compared to the other classifiers. Both deterministic spherical and elliptical classification have high variance. The statistical quantizers always achieves the best recognition rates, but GMM has a lot less variance than random, as expected since it models the actual error. For instance, GMM achieves a 71% recognition rate for 20 gestures, whereas deterministic spherical quantizer has a recognition rate of 33.8%. Statistical GMM quantization has a 98% recognition rate for 2 gestures.

Fig. 13: Gesture recognition rates for WP data set.
Fig. 14: Gesture recognition rate as a function of the number of gestures in the SW data set.

User-dependent and user-independent gestures To explore the sensitivity of recognition to the training data, we vary the number of users in the training data from 2 to 8. We compare with Costante et al. [14] and Liu et al. [11] which both perform this same experiment. We use six gestures from the Costante et al. SW data set: gestures 1, 3, 5, 7, 9, and 11. Costante et al. find that more users produces better accuracy, whereas Liu et al. find more personalized training (fewer users) works better. Table I presents accuracy for deterministic elliptical quantization as a function of users. In contrast, our recognition algorithm is not sensitive to the number of users and has high accuracy for both user-dependent (fewer users) and user-independent (more users) training.

Frequency-based personalization This section shows how our system easily incorporates additional sources of domain-specific information to improve accuracy. Suppose the gesture recognition application trains with 20 gestures from 8 people. In deployment, the gesture recognition application detects that the user makes 10 gestures with equal probability, but very rarely makes the other 10 gestures. We prototype this scenario, by expressing the user-specific distribution of the 20 gestures as a probability distribution in the programming framework in the classification code. At classification time, the runtime combines this distribution over the gestures with the HMM to improve accuracy. This figuration adds personalization to the deterministic elliptical quantization. Figure 15 shows how the distribution of gestures preformed by a specific user improves gesture recognition accuracy from 10 to 20% for each of the 10 gestures. Personalization could also be combined with statistical GMM.

Fig. 15: Classification accuracy of deterministic elliptical quantizer with personalization using .
Users Accuracy
2 82.71
4 85.09
6 84.15
8 82.08
TABLE I: Classification accuracy of deterministic elliptical quantizer gestures 1, 3, 5, 7, 9 and 11 from the SW data set as a function of the number of users training the HMM.

Balancing false positives and false negatives This experiment shows in more detail how the statistical quantization balances false positives with false negatives. In contrast, the deterministic elliptical quantizer always returns a classification with either a high probability (near 1) or low probability (near zero). Figure 18 shows a case study of classification of 10 gestures from the SW data set. The figure shows the recognition rate of a gesture whose recognition rate for deterministic elliptical quantizer is 0.90 and for statistical random quantizer and is 0.87. However, for the statistical random quantizer the balance between false positives and false negatives occurs at a higher threshold (near 0.5), which means that changing the conditional threshold of the classifier can decrease false negatives. However in the deterministic elliptical quantizer, the balance between false positives and false negatives happens at a lower conditional threshold, which means that the probability of false negatives is always higher in this classifier. Cost of statistical random quantization While the statistical random quantizer gives us the flexibility of higher precision or recall, it incurs more recognition time. We show that this overhead is low in absolute terms, but high compared to a classifier that does not explore multiple options. Figure 18 graphs how much time it takes to recognize different number of gestures with the deterministic elliptical quantization and statistical random quantization techniques. On average the statistical random quantizer is 16 times slower at recognizing 2–20 different gestures, taking 23 ms to recognize 20 gestures and 6.5 ms for two gestures. The statistical random quantizer uses .Pr() calls to invoke statistical hypothesis tests, and thus samples the computation many times. Figure 18 shows the default value for the .Pr() function. If we change the value of for the statistical test from 0.1 to 0.2, the time overhead reduces from 28 ms to 23 ms. If the system needs to be faster, statistical quantization trials are independent and could be performed in parallel. This additional overhead is very unlikely to degrade the user experience because in absolute terms, it is still much less than the 100 ms delay that is perceptible to humans [45].

Fig. 17: The default values for .Prob() inference calls.
Fig. 16: Balancing false positives and false negatives with the statistical random and GMM quantizer on 10 SW gestures.
Fig. 17: The default values for .Prob() inference calls.
Fig. 18: Time elapsed for classification in the statistical random and deterministic elliptical quantizer techniques.
Fig. 16: Balancing false positives and false negatives with the statistical random and GMM quantizer on 10 SW gestures.

X Conclusion

The promise of novel applications for sensing humans and machine learning is only realizable if, as a community, we help developers to use these tools correctly. This paper demonstrates that human gestures are very noisy and degrade the accuracy of machine learning models for gesture recognition. To help developers to more accurately deal with gesture noise, we introduce probabilistic quantization wherein gesture recognition finds the most likely sequence of hidden states given a distribution over observations rather than a single observation. We express this new approach using , a probabilistic programming system that automates inference over probabilistic models. We demonstrate how helps developers balance false positives with false negatives at gesture recognition time, instead of at gesture training time. Our new gesture recognition approach improves recall and precision over prior work on 25 gestures from 28 people.

References

  • [1] R. H. Liang and M. Ouhyoung. A real-time continuous gesture recognition system for sign language. In IEEE Automatic Face and Gesture Recognition. IEEE, 1998.
  • [2] T. Starner, J. Weaver, and A. Pentland. Real-time american sign language recognition using desk and wearable computer based video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1998.
  • [3] K. Hinckley. Synchronous gestures for multiple persons and computers. In Proceedings of the 16th annual ACM symposium on User interface software and technology, pages 149–158. ACM, 2003.
  • [4] Z. Ghahramani. An introduction to hidden Markov models and Bayesian networks. International Journal of Pattern Recognition and Artificial Intelligence, 2001.
  • [5] H-K Lee and J. H. Kim. An hmm-based threshold model approach for gesture recognition. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 1999.
  • [6] T. Schlomer, B. Poppinga, N. Henze, and S. Boll. Gesture recognition with a Wii controller. In ACM Conference on Tangible and Embedded Interaction, 2008.
  • [7] X. Zhang, X. Chen, W. Wang, J. Yang, V. Lantz, and K. Wang. Hand gesture recognition and virtual game control based on 3d accelerometer and emg sensors. In ACM conference on Intelligent User Interfaces, 2009.
  • [8] J. O. Wobbrock, A. D. Wilson, and Y. Li. Gestures without libraries, toolkits or training: A $1 recognizer for user interface prototypes. In ACM User Interface Software and Technology. ACM, 2007.
  • [9] J. Mäntyjärvi, J. Kela, P. Korpipää, and S. Kallio. Enabling fast and effortless customisation in accelerometer based gesture interaction. In Proceedings of the 3rd international conference on Mobile and ubiquitous multimedia. ACM, 2004.
  • [10] F. G. Hofmann, P. Heyer, and G. Hommel. Velocity profile based recognition of dynamic gestures with discrete hidden Markov models. In Gesture and Sign Language in Human-Computer Interaction. Springer, 1998.
  • [11] J. Liu, L. Zhong, J. Wickramasuriya, and V. Vasudevan. uwave: Accelerometer-based personalized gesture recognition and its applications. Pervasive and Mobile Computing, 2009.
  • [12] J. Bornholt, T. Mytkowicz, and K. S. McKinley. UncertainT: A first-order type for uncertain data. ASPLOS, 2014.
  • [13] T. Mytkowicz, J. Bornholt, a. Sampson, D. Z. Tootaghaj, and K. S. McKinley. UncertainT Open Source Project. https://github.com/klipto/Uncertainty/.
  • [14] G. Costante, L. Porzi, O. Lanz, P. Valigi, and E. Ricci.

    Personalizing a smartwatch-based gesture interface with transfer learning.

    In Signal Processing Conference (EUSIPCO), 2014 Proceedings of the 22nd European. IEEE, 2014.
  • [15] S. Mitra and T. Acharya. Gesture recognition: A survey. Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, 2007.
  • [16] A. Akl and S. Valaee. Accelerometer-based gesture recognition via dynamic-time warping, affinity propagation, & compressive sensing. In Acoustics Speech and Signal Processing (ICASSP), 2010 IEEE International Conference on. IEEE, 2010.
  • [17] G. Niezen and G. P. Hancke. Gesture recognition as ubiquitous input for mobile phones. In International Workshop on Devices that Alter Perception (DAP 2008), conjunction with Ubicomp. Citeseer, 2008.
  • [18] Z. Zhang. Microsoft kinect sensor and its effect. MultiMedia, IEEE, 2012.
  • [19] P. O. Kristensson, T. Nicholson, and A. Quigley. Continuous recognition of one-handed and two-handed gestures using 3d full-body motion tracking sensors. In Proceedings of the 2012 ACM international conference on Intelligent User Interfaces. ACM, 2012.
  • [20] D. Wilson and A. Wilson. Gesture recognition using the xwand. Technical Report CMURI-TR-04-57, CMU Robotics Institute, 2004.
  • [21] D. Mace, W. Gao, and A. Coskun. Accelerometer-based hand gesture recognition using feature weighted naïve Bayesian classifiers and dynamic time warping. In ACM Conference on Intelligent User Interfaces (companion). ACM, 2013.
  • [22] K. Murakami and H. Taguchi.

    Gesture recognition using recurrent neural networks.

    In Proceedings of the SIGCHI conference on Human factors in computing systems. ACM, 1991.
  • [23] VM Mantyla. Discrete hidden Markov models with application to isolated user-dependent hand gesture recognition. VTT publications, 2001.
  • [24] A. H. Ali, A. Atia, and M. Sami. A comparative study of user dependent and independent accelerometer-based gesture recognition algorithms. In Distributed, Ambient, and Pervasive Interactions. Springer, 2014.
  • [25] J. A. Hartigan and M. A. Wong.

    Algorithm as 136: A k-means clustering algorithm.

    Applied statistics, 1979.
  • [26] P. Paudyal, A. Banerjee, and S. KS Gupta. Sceptre: a pervasive, non-invasive, and programmable gesture recognition technology. In Proceedings of the 21st International Conference on Intelligent User Interfaces, pages 282–293. ACM, 2016.
  • [27] O. J. Woodman. An introduction to inertial navigation. University of Cambridge, Computer Laboratory, Tech. Rep. UCAMCL-TR-696, 14:15, 2007.
  • [28] A. B. C. Chatfield. Fundamentals of high accuracy inertial navigation, volume 174. Aiaa, 1997.
  • [29] CH Robotics Project. http://www.chrobotics.com/library/.
  • [30] E. M. Foxlin. Motion tracking system, 2008. US Patent 7,395,181.
  • [31] J. B. Kruskal and M. Wish. Multidimensional scaling, volume 11. Sage, 1978.
  • [32] L. E. Baum and T. Petrie. Statistical inference for probabilistic functions of finite state Markov chains. The annals of mathematical statistics, 1966.
  • [33] L. E. Baum, T. Petrie, G. Soules, and N. Weiss. A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains. The annals of mathematical statistics, 1970.
  • [34] S. Russell and P. Norvig. Artificial intelligence: A modern approach. Prentice Hall, 1995.
  • [35] N. Seshadri and C. Sundberg. List viterbi decoding algorithms with applications. IEEE Transactions on Communications, 42(234):313–323, 1994.
  • [36] D. G. Lowe. Object recognition from local scale-invariant features. In Computer vision, 1999. The proceedings of the seventh IEEE international conference on, volume 2. IEEE, 1999.
  • [37] D. G. Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60, 2004.
  • [38] J. C Mitchell, A. Ramanathan, A. Scedrov, and V. Teague. A probabilistic polynomial-time process calculus for the analysis of cryptographic protocols. Theoretical Computer Science, 2006.
  • [39] P. Xie, J. H. Li, X. Ou, P. Liu, and R. Levy. Using Bayesian networks for cyber security analysis. In Dependable Systems and Networks (DSN), 2010 IEEE/IFIP International Conference on. IEEE, 2010.
  • [40] L. Ngo and P. Haddawy. Answering queries from context-sensitive probabilistic knowledge bases. Theoretical Computer Science, 1997.
  • [41] O. Kiselyov and C. Shan. Embedded probabilistic programming. In Domain-Specific Languages. Springer, 2009.
  • [42] A. Sampson, P. Panchekha, T. Mytkowicz, K. S. McKinley, D. Grossman, and L. Ceze. Expressing and verifying probabilistic assertions. In ACM Conference on Programming Language Design and Implementation (PLDI). ACM, 2014.
  • [43] C. Nandi et al. Debugging probabilistic programs. In Proceedings of the 1st ACM SIGPLAN International Workshop on Machine Learning and Programming Languages. ACM, 2017.
  • [44] E. Meijer, B. Beckman, and G. Bierman. Linq: Reconciling object, relations and xml in the .net framework. In Proceedings of the 2006 ACM SIGMOD International Conference on Management of Data, SIGMOD ’06, pages 706–706, New York, NY, USA, 2006. ACM.
  • [45] Stack Overflow. Can a human eye perceive a 10 milliseconds latency in image load time? http://stackoverflow.com/q/7882713/39182/, accessed in April, 2016.