Page 163 - 2024-Vol20-Issue2

159 | Murad & Alasadi

TABLE II. VII. DEEP LEARNING APPROACHES
A COMPARISON OF MACHINE LEARNING ALGORITHMS
Computer Vision (CV) applications have evolved from human
FOR HAND GESTURE RECOGNITION. activity recognition to speech recognition, image classifica-
tion, and labeling [62]. Deep learning, a machine learning
Algorithm Pros Cons technology, has emerged as a successful solution since 2010.
Artificial Can learn complex Can be computation- Researchers have transitioned from traditional handcrafted
Neural features and pat- ally expensive, diffi- features to data-driven algorithms since 2010. Various learned-
Networks terns, with high cult to train based feature methods, such as genetic programming and
accuracy dictionary-based approaches, have been used for visual recog-
Hidden Can handle sequen- Requires training, nition tasks. Still, deep learning has significantly influenced
Markov tial data, with high can be difficult to computer vision in recent years.
Models accuracy define the number Vision-based recognition algorithms include handcrafted fea-
of states tures and learned-based features [63]. Handcrafted features
Fuzzy Flexible grouping Can be computation- resolve CV problems by collecting correct features from data.
Clustering of data can handle ally expensive The pipeline typically includes four major stages; Section 3.1
Algo- noisy data provides further details.
rithms Sensitive to the Unlike handcrafted methods, deep learning uses convolution
K-nearest Simple to im- choice of k, can neural networks to learn features automatically during train-
Neighbor be computationally ing. This approach interprets problems in terms of concep-
plement and expensive for large tual hierarchy, with lower network concepts encoding simple
Support datasets representations and high-level layers constructing abstract
Vector understand, can Can be computation- concepts. This hierarchical learning process allows for the
Machines ally expensive for complete elimination of handcrafted feature extraction and
Dynamic handle numerical large datasets allows convolution neural networks to function as end-to-end
Time Can be computation- learners.
Warping and categorical data ally expensive for Traditional algorithms like Support Vector Machine, Random
large datasets Forest, and Hidden Markow Model rely heavily on data rep-
High accuracy, ro- resentation, but handcrafted features often cause information
loss. Deep learning algorithms have shown impressive re-
bust to noise sults in CV challenges but require large amounts of data for
accurate learning. Deep neural network training and param-
Can handle sequen- eterization also require significant computational resources
tial data with vary- and experimentation time [64].
ing time scales, high
accuracy VIII. HAND GESTURE RECOGNITION
APPLICATIONS
classifiers has been suggested to break multi-class problems
into multiple binary problems. Hand Gesture Recognition is crucial in various applications,
including human-computer interaction, sign language recog-
F. Dynamic Time Warping nition, virtual and augmented reality, and robotics. Characters
Dynamic Time Warping (DTW) algorithm is widely used use hand gestures in future films, highlighting how interac-
for finding the best alignment between two signals [60]. It tion systems can influence our interactions with computers if
computes the distance between point pairs and their related current systems do not allow such freedom.
feature values, computes a cumulative distance matrix, and
obtains the smallest costly path. This path reflects the perfect A. Virtual and Augmented Reality Technologies
warp, reducing feature distance among synchronized points. Virtual Reality (VR) is a digital technology that mimics a hu-
DTW is used in data mining, speech recognition, and motion man’s presence in a virtual setting by creating sounds, images,
recognition. Prior work has mainly focused on algorithm ac- and feelings using headsets [65]. AR, or augmented reality,
celeration, constraint analysis, algorithm approximation, and improves the real world, while VR completely replaces it with
lower bounding methods. an artificial one. American teen Palmer Luckey built a VR
Derivative DTW (DDTW) is a form of DTW that measures headset prototype in 2010, which evolved into the Oculus Rift.
distances between first-order derivatives of points, focus-
ing on shape characteristics [61]. Most of the work is one-
dimensional.
Table II illustrates a concise table of machine learning algo-
rithms for hand gesture recognition.