Comparative evaluation of static gesture recognition techniques based on nearest neighbor, neural networks and support vector machines

Abstract

It is a common behavior for human beings to use gestures as a means of expression, as a complement to speaking, or as a self-contained communication mode. In the field of Human-Computer Interaction, this behavior can be adopted to build alternative interfaces, aiming to ease the relationship between the human element and the computational element. Currently, various gesture recognition techniques are described in the technical literature; however, the validation studies of these techniques are usually performed isolatedly, which complicates comparisons between them. To reduce this gap, this work presents a comparison between three well-established techniques for static gesture recognition, using Nearest Neighbor, Neural Networks, and Support Vector Machines as classifiers. These classifiers evaluate a common dataset, acquired from an instrumented glove, and generate results for precision and performance measurements. The results obtained show that the classifier implemented as a Support Vector Machine presented the best generalization, with the highest recognition rate. In terms of performance, all methods presented evaluation times fast enough to be used interactively. Finally, this work identifies and discusses a set of relevant criteria that must be observed for the training and evaluation steps, and its relation to the final results.

Max-pooling convolutional neural networks for vision-based hand gesture recognition

“…Since SVMs [29] have gained much attention in recent times due to their powerful generalization capabilities as gesture classifiers [16], [18] we evaluate different feature learning schemes using SVMs. The following approaches are evaluated in this paper using our dataset: (i) The authors in [30], [31], [32] use Hu Invariant Moments for feature learning from images of different objects and gestures; (ii) Unsupervised feature learning is applied by authors in [33] using the Spatial Pyramid (generally referred to as Bag of Features or Bag of Words (BoW)) a combination of SIFT and k-means; (iii) Shape properties of objects such as roundness, form factor, compactness, eccentricity, perimeter, solidity etc are used by the authors in [31], [34]; (iv) Skeletonization has been proposed by the authors in [35], [36] for gesture recognition tasks, such as the counting the number of fingers; (v) Pyramid of Histogram Oriented Gradients (PHOG) [37], a variant of the famous HOG descriptor [38], gained popularity for its vectorized HOG feature learning approach; (vi) The Fast Fourier Transform (FFT) has been used by the authors in [39] to represent the shape of the hand contour in images using the spatial domain; (vii) CNNs called Tiled CNNs [40] are supervised feature learners and classifiers able to learn complex invariances such as scale and rotational invariance.…”

“…Recently, different research efforts on 2D appearance model-based methods for gesture recognition have emerged [9], [10], [11], [12], [13], [14], [15], amongst which supervised and unsupervised learning techniques such as Neural Networks (NNs), Support Vector Machine (SVMs) and NearestNeighbor [16], [17], [18] classifiers have gained familiarity. However, feature learning is not a part of such classification schemes and needs to be performed separately to compute features such as edges, gradients, pixel intensities and object shape.…”

Recognizing human activities based on head movement trajectories

“…But the space of HAR is limited into a small part of the environment, where the HAR sensor has been placed. Obviously, absolute anchoring approaches are principally vision-based methods and can be further categorized by the function of modeling variations in time: direct classification, which classifies image features without using information about the time factor and HAR is usually performed directly for each frame individually such as in [12], [13], [14] and [15]. Furthermore, temporal state-space methods where temporal data appears as a particular dimension and where every observation is equivalent to an image representation in given a time such as in [16], [17], [18] and [19].…”