2. Motivation
• Build a classification model that identifies
spherical and tip grasps from surface
electromyography (sEMG) data
3. Types of Grasps
Spherical (Power grip) Tip (Precision grip)
Image: https://www.studyblue.com/notes/note/n/grasps-
updated-for-human-occ/deck/4455956
Image:
https://www.studyblue.com/notes/note/n/grasps-
updated-for-human-occ/deck/4455956
4. Dataset
• Healthy, male subject performs each grasp 100 times daily for three
days
• 600 instances (tip and spherical grasp data combined)
• Dataset was originally band-passed filtered to eliminate line
artifacts
Reference: C. Sapsanis. 'Recognition of basic hand movements using electromyography'. 2013.
5. Building a Classifier
1. Specify network model
2. Data pre-processing and Feature Extraction
3. Create Training and Test Sets
4. Parameterize the network
5. Validate and query model results
6. Hidden Markov
Model
• Simplest form of a
Dynamic Bayesian
Network
• Utilizes local conditional
probability densities to
compactly specify a
joint distribution
• Latent variables are
discrete random
variables
• Observed variables are
continuous random
variables
N
n
N
n
nnnnNN goPggPgPggooP
2 1
1111 )/()]/()[(),...,,,...,(
7. Model Variables and Feature
Extraction
Feature Extraction:
𝐼𝐸𝑀𝐺 =
𝑛=1
𝑁
𝑥 𝑛
used for observations denoted 𝑂 𝑛 in HMM
𝐺 𝑛 is a binary random variable taking values in the set {0,1}
where
0 – spherical grasp
1 – tip grasp
Reference: A. Phinyomark, C. Limsakul, and P. Phukpataranont. “A Novel Feature Extraction for Robust EMG Pattern
Recognition.” Journal of Computing. Volume 1. Issue 1, December 2009.
8. Training and Test sets
• Randomly sampled dataset
• 90 percent of data for training
• Remaining 10 percent used for testing
9. Parameterization
• Goal: fit the joint probability distribution to
the data, i.e. find the parameters
• Bayesian Logic (BLOG) computes posterior
distributions based on specified priors and
observation values using sampling algorithms
• Particle filter chosen to run the HMM
10. Computing Posterior Distributions
Specified priors in BLOG:
P(𝐺1) – probability mass function (pmf)
𝑃(𝐺 𝑛|𝐺 𝑛−1) – conditional probability table
𝑃(𝑂𝑛|𝐺 𝑛) – Gaussian distribution
=
1
𝜎 2𝜋
𝑒− (𝑥−𝜇 𝑖)2
2𝜎2 , i = 0, 1
𝑃(𝐺 𝑛|𝑂 𝑛) ∝ 𝑃(𝑂𝑛|𝐺 𝑛) 𝑃(𝐺 𝑛 )
12. Conclusion
• Current model predicts spherical grasps more
accurately than tip grasps
• Beneficial for picking up large objects as spherical
grasps are more effective for this task as
compared to tip grasps
• Improvement can be made to increase accuracy
– 1. Include additional features to feature vector and try
various combinations of vectors
– 2. Stratify the random sample when creating training
and test sets to ensure data from both groups is
represented