American Sign Language Hand Gesture Recognition (2023)

American Sign Language Hand Gesture Recognition (1)

This project and blog was a joint effort by Rawini Dias, LaShay Fontenot, Katie Grant, Chris Henson, and Shivank Sood.
Please visit our
Github repository for the project implementation code.

While there are new and accessible technologies emerging to help those with hearing disabilities, there is still plenty of work to be done. For example, advancements in machine learning algorithms could help the deaf and hard-of-hearing even further by offering ways to better communicate using computer vision applications. Our project aims to do just that.

We sought to create a system that is capable of identifying American Sign Language (ASL) hand gestures. Since ASL has both static and dynamic hand gestures, we needed to build a system that can identify both types of gestures. This article will detail the phases of our project.

Goal: Build a system that can correctly identify American Sign Language signs that corresponds to the hand gestures

Method: The static sign language data for our project was in the form of images. We trained a Convolutional Neural Network (CNN) to identify the signs represented by each of these images. The dynamic sign language dataset we used was collected by a LeapMotion Controller (LMC) and was in the form of (x, y, z) coordinates of each joint of each hand collected every few milliseconds. We feature engineered this data to get useful relative motion data which was then trained on classical classification models to identify the specific sign pertaining to each LMC input.

Applications: Our proposed system will help the deaf and hard-of-hearing communicate better with members of the community. For example, there have been incidents where those who are deaf have had trouble communicating with first responders when in need. Although responders may receive training on the basics of ASL, it is unrealistic to expect everyone to become fully fluent in sign language. Down the line, advancements like these in computer recognition could aid a first responder in understanding and helping those that are unable to communicate through speech.

Another application is to enable the deaf and hard-of-hearing equal access to video consultations, whether in a professional context or while trying to communicate with their healthcare providers via telehealth. Instead of using basic chat, these advancements would allow the hearing-impaired access to effective video communication.

Performance: The proposed model for the still images is able to identify the static signs with an accuracy of 94.33%. Based on our analysis of the dynamic signs, we realized the need to identify if the sign is a one-handed or two-handed sign first, and then identify the sign itself. The final model we propose for the dynamic signs is capable of identifying the one-handed signs with an accuracy of 88.9% and the two-handed signs with an accuracy of 79.0%.

Data Collection & Pre-Processing

Using the Sign Language MNIST dataset from Kaggle, we evaluated models to classify hand gestures for each letter of the alphabet. Due to the motion involved in the letters J and Z, these letters were not included in the dataset. However, the data includes approximately 35,000 28x28 pixel images of the remaining 24 letters of the alphabet. Similar to the original MNIST hand drawn images, the data contains an array of grayscale values for the 784 pixels in each image. One of these images is shown below.

American Sign Language Hand Gesture Recognition (2)


We used a Convolutional Neural Network, or CNN, model to classify the static images in our first dataset. Our first goal when building the neural network was to define our input layer. A 28x28 image contains 784 pixels each represented by a grayscale value ranging from 0 (black) to 1 (white). By converting each image to a series of numbers, we transform the data into a format the computer can read.

(Video) Easy Hand Sign Detection | American Sign Language ASL | Computer Vision

Once the input layer has been prepared, it can be processed by the neural network’s hidden layers. The architecture of our neural network can be seen below.

The first hidden layer is composed of several nodes each of which takes a weighted sum of the 784 input values. The weighted sum of inputs is then input into an activation function. For our network, we used a rectified linear unit, or ReLU.

American Sign Language Hand Gesture Recognition (4)

The above graph shows that the ReLU will output 0 when the input is negative, but will not change the input otherwise. The outputs from the ReLU will serve as the inputs to the next hidden layer in the network.

To better understand how each hidden layer transforms the data, we can visualize each layer’s outputs. Our first layer had 32 channels, so the process described above was repeated 32 times. This allows the network to capture several features in each image. If we input the image depicting the letter “C” that was shown previously, we obtain the following set of 32 outputs.

American Sign Language Hand Gesture Recognition (5)

Here, we see how each channel transforms the image a little differently. Based on these images, it appears the network is extracting information about the edges and general shape of the person’s hand. As the data continues to move through the hidden layers, the neural network attempts to extract more abstract features. Below are the outputs of the fourth hidden layer. These images are much less interpretable to the human eye, but will be very useful to the network as it attempts to classify the image into 1 of 24 potential classes.

American Sign Language Hand Gesture Recognition (6)

Once the data has passed through the Convolution and MaxPool layers of the neural network, it enters the Flatten and Dense layers. These layers are responsible for reducing the data to one dimension and identifying an image’s class.

After the CNN’s architecture was defined, we attempted to optimize the model’s performance by selecting an appropriate value for the number of epochs. Earlier we mentioned that each node takes a weighted sum of its inputs. The weights applied to each input are learned through the training process and updated with each epoch. An epoch is a single pass through all of the training data. On the first epoch, the neural network estimates a value for each weight. For each subsequent epoch, the neural network updates these weights with values that reduce overall loss. Generally, more epochs result in more accurate classifiers; however, more epochs also produce more complex models. Using a validation set, we determined that 10 epochs provided us with the best balance between accuracy and complexity.

(Video) Hand Gesture Recognition (American Sign Language) using Python


The training and validation datasets used to build and optimize the model contained 80% of the original data. The remaining 20% (~7,000 samples) was reserved for model testing. When this test data was input to the model, it achieved 94.33% accuracy. To further understand the strengths and weaknesses of this model, we created a confusion matrix.

American Sign Language Hand Gesture Recognition (7)

From the confusion matrix, we see that the two signs most commonly confused are the letters “M” and “S”. Images of each of these signs are shown below.

American Sign Language Hand Gesture Recognition (8)
American Sign Language Hand Gesture Recognition (9)

Based on these images, it is easy to understand why our neural network has trouble distinguishing between these two signs. In future work, we will use images with higher resolution that allow for more intricate details to be extracted from the images. Hopefully, this will further improve the accuracy of our model. An additional limitation of this model was its inability to recognize moving signs, such as the letters “J” and “Z”. In the next section, we explore a data source that is better equipped to recognize dynamic signs.

Understanding the Data

The second phase of our project will focus on dynamic signs (i.e. moving signs). This dataset consists of 25 subjects each performing the same 60 ASL signs with both their left and right hands using a LeapMotion Controller (LMC). Therefore, this dataset has 60 different ASL signs (or class labels) that we are trying to accurately predict.

The LMC device records the position of the fingers, joints, palm, wrist, and arm every 0.04 seconds. In other words, the LMC acquires spatial coordinates of the skeleton joints of the hands and how these coordinates vary with time. Our second dataset is made up of these coordinate points.

Let us try to understand the nature of this data in more detail using Figure 9.

American Sign Language Hand Gesture Recognition (10)

The metacarpals, proximal, intermediate, and distal bones refer to the four different bones of an anatomical finger. The Leap Motion dataset gives us the (x, y, z) coordinates of each of these bones in each finger every 0.04 seconds for the duration of the sign. It also gives us the coordinates for the palm, wrist, and arm. Altogether, these coordinate points as a function of time provides discriminating information that can be used to identify the type of hand gesture (or ASL sign).

Feature Engineering

Realizing that we needed our dataframes for each test subject to be comparable, we first transformed each dataframe by taking the difference of each successive row of coordinates, giving the distance that each part of the hand (x, y, and z coordinates) traveled in between each measurement that the Leap Motion device recorded. This captured movement in intervals of approximately .04 seconds for the 54 parts of the hand identified during the motion capture. Each of these transformed data frames consisted of anywhere from 398–1203 time intervals, each with 162 columns of coordinate data.

After this differencing, we next sought to derive features that captured information about the movement of the hand during the given time interval. We decided to take the mean and standard deviation of each of these columns. While this may see relatively simplistic, we found this a computationally cheap way to capture information.

Consider what the mean of each of these differenced columns represents. This is the average distance traveled by each part of the hand in each time interval. Likewise, taking the standard deviation of each of these columns represents the variation in this displacement. In other words, this serves as a proxy for velocity. (Surprisingly, adding the actual calculation for velocity at each point actually reduced accuracy!)

In addition to these two sets of features, we also experimented with identifying pairs of hand coordinates with strong correlation and using the polynomial weights as features for classification. While this produced incremental gains in accuracy in a reduced data set (10 classes as compared to all 60) this did not scale well to the full set of signs.

(Video) Sign Language Recognition Using Hand Gestures

Another idea we had was to use the angles the fingers formed. We calculated the internal angles of the joints between distal and intermediate bones and the internal angles of the joints between intermediate and proximal bones. However, adding these angles to the previously derived mean and standard deviation features for each joint indicated multicollinearity in the independent variables. Using only the angle features derived resulted in acceptable classification accuracy on the reduced data set of 10 classes, but it did not extend well to the dataset with all 60 classes.


Initially our team began model selection by looking at the subset of our data that consisted of numeric signs zero through ten, developing the features described above. What we realized is that this subset of signs could be easily distinguished by the fact that each of them only utilizes a single hand to complete the sign. Realizing this meant that the still left hand was only contributing noise to the data set, we removed all coordinates originating from the left hand and saw a significant gain in classification accuracy.

Now wanting to extend this to our full data set, we used both hands, again with the above features, and noted a significant decline in accuracy. In an attempt to identify where our model was unable to distinguish between different signs, we found that a better understanding of sign language would inform our model pipeline.

We first attempted to conditionally identify which signs utilized only one hand, with the intent of dividing our data set into two groups. Through both manually examining the signs and developing thresholds for our feature means, we split the data set into 22 “two-handed” signs and 38 “one-handed” signs by identifying data frames whose left handed attributes appeared to be still, as measured by the sum of mean absolute deviation in left-handed coordinates. This distinction however, is often far from clear.

As members of our team are far from fluent in American Sign Language, we had to do some research to understand more about the signs in our dataset. We set out to determine how many, and which hands were involved in signing each word. Fortunately for us, the website Signing Savvy offers an ASL dictionary complete with videos of the various ways to sign each word. For example, the word ‘bug’ can be signed in two different ways depending on the manner in which the word is being used. Both of these signs used only the right hand when signing. On the other hand, the word ‘cost’ involves both hands, but only one hand is in motion (video). Using this website, we were able to understand the application-based differences of the signs in our dataset. Of the 60 words in our dataset, there were 9 that could be signed using either both hands, or just the right hand (Car Drive, Come, Cost, Finish, Go, Happy, Hurt, Small, When).

Regardless of the difficulties associated with splitting our data set between one-handed and two-handed signs, we found that this methodology significantly increased our accuracy for the complete data set.

Model Selection

After creating the previously mentioned features, we experimented with several classifiers. Below are two boxplots, showing the performance of various models on our one-handed and two-handed signs. Ultimately, we decided that linear discriminant analysis (LDA) had superior performance. Below are our accuracy results, run for 100 test/training splits:


The plots in Figure 12 show the results of our model run one hundred times using LDA, each iteration taking a different randomly selected training/test split stratified by our classification label. Two pieces of information immediately stand out. First, the tactic of separating one vs. two handed signs is very useful. Second, our model accuracy has a relatively high standard deviation with regard to classification accuracy given a random set of training data. Considering the small sample size of 25 test subjects however, this should not be much of a surprise.

American Sign Language Hand Gesture Recognition (13)

The above plot demonstrates that our two-handed signs are systematically misclassified more often than our one-handed signs. The natural question is to identify the particular signs that may be problematic for our model. The most commonly confused signs are shown in the table below:

(Video) Video of gesture recognition and ASL translation.

American Sign Language Hand Gesture Recognition (14)

Table 1 shows the signs that were misclassified with each other, i.e. the sign for ‘come’ was misclassified as either ‘big’ or ‘with’. Similarly, the sign for ‘red’ was misidentified as ‘cry’ in our dataset. This was to be expected as these pairs of signs were very similar in motion.

Furthermore, the overall model misclassifications can be categorized into two groups.

  1. Signs that are “over predicted”. In other words, signs that are predicted when the actual sign is something else. This is the equivalent to false positives in a two-class problem.
  2. Signs that are “under predicted”. In other words, signs that are not predicted when they should be. This is the equivalent to false negatives in a two-class problem.

Below are two charts that identify these cases:

American Sign Language Hand Gesture Recognition (15)
American Sign Language Hand Gesture Recognition (16)

For example, the sign “Cold” (link to video) often fails to be predicted. Intuitively, what we are seeing is that because this sign involves both hands moving through a very small range of motion that it is very easy for our model to predict this in place of another sign.

In order to further visualize how these signs behave in our classification model, we generated ROC plots for each sign based on a “One Versus Rest” Classification (still using LDA), where for each individual sign we treat our model as working with a two class model (for instance “Cold” versus “not Cold”). Below we see the results for one of the most problematic signs:

American Sign Language Hand Gesture Recognition (17)

While this analysis sets a solid baseline for American Sign Language recognition, more work needs to be done to apply this concept in real-time. This would require further work with the LeapMotion API to enable real-time generation of data, feeding through the model, and identification of the word and/or numbers. This would also require the model to be able to handle more than the 60 class labels it currently deals with.

In conclusion, we see this application having real potential in improving the lives of the hearing-impaired and as such it would be a worthy goal to continue development.

(Video) 35 Gesture Recognition Using Sign Language MNIST




What is sign language and hand gesture? ›

Hand gestures are used as a way for people to express thoughts and feelings, it serves to reinforce information delivered in our daily conversation. Sign language is a structured form of hand gestures involving visual motions and signs, which are used as a communication system.

What is gesture recognition example? ›

For example, imagine being able to check your home security camera as you drive home by simply making a hand gesture. Gestures could also be coupled with telematics systems, allowing the vehicle to provide information about nearby landmarks if it recognizes that an occupant is pointing at it.

How many gestures are there in American Sign Language? ›

ASL possesses a set of 26 signs known as the American manual alphabet, which can be used to spell out words from the English language. Such signs make use of the 19 handshapes of ASL.

What are the limitations of sign language recognition? ›

The disadvantages are that they are costly and are difficult to be used commercially. Classification methods are also varying from researchers. Researchers tend to develop their own concept, based on known methods, to give better result in recognizing the sign language.

What does 🤟 mean in sign language? ›

What does 🤟 I Love You Gesture emoji mean? A universal emoji! Or … is it? The love-you gesture or I love you hand sign emoji is the American Sign Language gesture for “I love you,” showing a hand with a raised index finger and pinky (little) finger and an extended thumb.

What is the difference between sign language and gesture? ›

A sign language is a formal, agreed-upon set of movements used to replace spoken language. Gestures in general—such as rolling the eyes, shrugging, or flipping the bird—are less formal and don't have a systematized set of meanings.

How do you identify hand gestures? ›

Recognition of Hand Gestures. When the fingers are detected and recognized, the hand gesture can be recognized using a simple rule classifier. In the rule classifier, the hand gesture is predicted according to the number and content of fingers detected. The content of the fingers means what fingers are detected.

What is meaning of gesture recognition? ›

Gesture recognition is a type of perceptual computing user interface that allows computers to capture and interpret human gestures as commands. The general definition of gesture recognition is the ability of a computer to understand gestures and execute commands based on those gestures.

What is hand gesture control? ›

Gesture control is the ability to recognize and interpret movements of the human body in order to interact with and control a computer system without direct physical contact.

What does ✌ mean in sign language? ›

The victory hand emoji, ✌️, is a representation of the peace sign.

What does ✊ mean in sign language? ›

Emoji Meaning

A person gesturing with their index finger between ear and mouth, used as a deaf sign in American Sign Language (ASL) and a number of other global sign languages. See also Ear With Hearing Aid. Deaf Person was approved as part of Unicode 12.0 in 2019 and added to Emoji 12.0 in 2019.

What does rubbing your chin mean in sign language? ›

Thank You:

Touch the chin or lips with the fingertips of one flat hand, then move the hand forward until the palm is facing up. The hand moves out and down. This sign is similar to the gesture of kissing one's hand and extending the hand towards someone else.

Which algorithm is used for sign language recognition? ›

We conclude that SVM+HoG and Convolutional Neural Networks can be used as classification algorithms for sign language recognition.

Is there an app that reads sign language? ›

About this app

Led by Hugo, the world's most lovable 3D interpreter, the Hand Talk app automatically translates text and audio to American Sign Language (ASL) [Beta] and Brazilian Sign Language (Libras) through artificial Intelligence.

Why do we need ASL detection? ›

Deaf and Mute people use hand gesture sign language to communicate, hence normal people face problems in recognizing their language by signs made. Hence there is a need for systems that recognize the different signs and conveys the information to normal people.

How do you say kiss me in sign language? ›

To sign kiss, start by extending your fingers and holding them together. Then touch your mouth, followed by your cheekbone. It is like you showing someone how to give a cheek kiss. They take their lips and touch them on your cheek.

What is the most common sign language? ›

ASL, short for American Sign Language, is the sign language most commonly used by the Deaf and Hard of Hearing people in the United States. Approximately more than a half-million people throughout the US (1) use ASL to communicate as their native language.

What is I Love U in sign language? ›

The sign for “I love you” is a combination of the fingerspelled letter I, L and Y. Your thumb and index finger together form an L, while your little finger forms an I. In addition, your thumb and little finger is expressing a Y. So if you combine all three handshapes, you get I-L-Y for I love you.

Does sign language count as a gesture? ›

The gestures or symbols in sign language are organized in a linguistic way. Each individual gesture is called a sign. Each sign has three distinct parts: the handshape, the position of the hands, and the movement of the hands. American Sign Language (ASL) is the most commonly used sign language in the United States.

Is sign language verbal or nonverbal? ›

Sign language is a non-verbal language that Deaf persons exclusively count on to connect with their social environment. It is based on visual cues through the hands, eyes, face, mouth, and body. The gestures or symbols in sign language are organised in a linguistic way.

Are gestures considered language? ›

Researchers have discovered that gesture is an integral part of language—it forms a unified system with speech and, as such, plays a role in processing and learning language and other cognitive skills.

What is sign language example? ›

Sign language is defined as a way to communicate using hand gestures and symbols for words or letters of the alphabet, often used by those who are hard-of-hearing. An example of sign language is the means of communicating used by Helen Keller.

What is the meaning of hand signs? ›

Definition of hand signal

: a movement of a person's hands that means something They communicated with each other by using hand signals.

What is sign language in health and social care? ›

Sign language is a way of communicating visually, using hand gestures, facial expressions and body language.

Why is sign language used? ›

Sign language is used mainly by people who are Deaf or have hearing impairments.

What is the most used sign language? ›

ASL, short for American Sign Language, is the sign language most commonly used by the Deaf and Hard of Hearing people in the United States. Approximately more than a half-million people throughout the US (1) use ASL to communicate as their native language.

What are the 3 types of sign language? ›

Not a Universal Language

Interestingly, most countries that share the same spoken language do not necessarily have the same sign language as each other. English for example, has three varieties: American Sign Language (ASL), British Sign Language (BSL) and Australian Sign Language (Auslan).

What is I Love U in sign language? ›

The sign for “I love you” is a combination of the fingerspelled letter I, L and Y. Your thumb and index finger together form an L, while your little finger forms an I. In addition, your thumb and little finger is expressing a Y. So if you combine all three handshapes, you get I-L-Y for I love you.

What are the 4 types of gestures? ›

McNeill (1992) proposes a general classification of four types of hand gestures: beat, deictic, iconic and metaphoric.

What does 2 fingers sideways mean? ›

What does two fingers held sideways mean? Holding up two fingers, turned sideways indicates “peace" or “goodbye.” This sign is generally called “deuces" or “peace.”

What does throwing up 3 fingers mean? ›

Three-finger salute (pro-democracy), a gesture originally from the Hunger Games books and films and later used in protests in Myanmar and Thailand. Three-finger salute, a jocular term for the three-key command Control-Alt-Delete.

What is the meaning of middle finger in sign language? ›

The middle finger can also be substituted for the standard hand shape of an otherwise innocuous sign. The location, orientation, and movement of the original sign are retained, with the only difference being inclusion of the middle finger. This may be done for the sake of emphasis, The non-profane form of “Understand.”

What is sign language in simple words? ›

Sign language is manual communication commonly used by people who are deaf. Sign language is not universal; people who are deaf from different countries speak different sign languages. The gestures or symbols in sign language are organized in a linguistic way.

Who invented sign language? ›

The first person credited with the creation of a formal sign language for the hearing impaired was Pedro Ponce de León, a 16th-century Spanish Benedictine monk. His idea to use sign language was not a completely new idea.

What are 5 interesting facts about sign language? ›

Five Interesting Facts Most People Don't Know About Sign Language
  • It's the fourth most used language in the UK. ...
  • Different countries have their own versions of sign language. ...
  • Sign language uses more than just hand gestures. ...
  • Many deaf people have 'name signs' ...
  • Sign language isn't as difficult to learn as it looks.
Sep 23, 2021

What do you call a sign language person? ›

A signer is a person who can communicate conversationally with people who are deaf or hard of hearing. An interpreter is a person who is not only bilingual but has also received specialized training and credentials to develop the skills and expertise needed to mediate meanings across languages and cultures.

Do all deaf people use American Sign Language? ›

ASL is the primary language for deaf and hard of hearing people to communicate, but not all deaf and hard of hearing people use it.


1. Hand gesture recognition using python and opencv
(Sadaival Singh)
2. Real Time Sign Language Detection with Tensorflow Object Detection and Python | Deep Learning SSD
(Nicholas Renotte)
3. Custom Hand Gesture Recognition with Hand Landmarks Using Google’s Mediapipe + OpenCV in Python
(Ivan Goncharov)
4. American Sign Language : Hand Gestures to Words
5. Sign Language Detection using ACTION RECOGNITION with Python | LSTM Deep Learning Model
(Nicholas Renotte)
6. Sign Language Recognition using Machine Learning | Sign Language Detection | Data Magic
(Data Magic (by Sunny Kusawa))
Top Articles
Latest Posts
Article information

Author: Foster Heidenreich CPA

Last Updated: 10/24/2022

Views: 6331

Rating: 4.6 / 5 (56 voted)

Reviews: 87% of readers found this page helpful

Author information

Name: Foster Heidenreich CPA

Birthday: 1995-01-14

Address: 55021 Usha Garden, North Larisa, DE 19209

Phone: +6812240846623

Job: Corporate Healthcare Strategist

Hobby: Singing, Listening to music, Rafting, LARPing, Gardening, Quilting, Rappelling

Introduction: My name is Foster Heidenreich CPA, I am a delightful, quaint, glorious, quaint, faithful, enchanting, fine person who loves writing and wants to share my knowledge and understanding with you.