Tea Marasovic

A novel feature descriptor for gesture classification using smartphone accelerometers

Since gestures are a natural form of human expression, gesture-based interfaces can serve as an alternative interaction modality with numerous aspects to be utilized in human computer interaction. In this paper, we address the issue of finding a compact but effective set of features for a robust gesture recognition, using a single 3-axis accelerometer. A novel feature extraction scheme, that allows the gesture form to be clearly discriminated, is proposed. Fuzzy k-Nearest Neighbour classifier is used for recognition of gestures in transformed feature space. The experiments, conducted on an custom gesture vocabulary, reveal that Histogram of Direction (HoD) descriptor, in conjunction with statistical features, produces a highly competitive performance, in terms of recognition accuracy.

Accelerometer based gesture recognition system using distance metric learning for nearest neighbour classification

The need to improve communication between humans and computers has been motivation for defining new communication models, and accordingly, new ways of interacting with machines. In many applications today, user interaction is moving away from traditional keyboards and mouses and is becoming much more physical, pervasive and intuitive. This paper examines hand gestures as an alternative or supplementary input modality for mobile devices. A new gesture recognition system based on the use of acceleration sensor, that is nowadays being featured in a growing number of consumer electronic devices, is presented. Accelerometer sensor readings can be used for detection of hand movements and their classification into previously trained gestures. The proposed system utilizes Mahalanobis distance metric learning to improve the accuracy of nearest neighbour classification. In the approach we adopted, the objective function for metric learning is convex and, therefore, the required optimization can be cast as an instance of semidefinite programming. The experiments, carried out to evaluate system performance, demonstrate its efficacy.

Performance of Compressive Sensing Image Reconstruction for Search and Rescue

In this letter, a system combining compressive sensing (CS)-based image reconstruction and object detection algorithm is introduced. The use of CS is a promising approach for search-and-rescue applications, since it highly reduces the amount of data that needs to be transmitted. However, the high-quality reconstruction of such images is a challenging task due to the complexity of structures and the number of tiny details, possibly being the objects of interest. Hence, the performance of image reconstruction is evaluated in terms of the missing data amount and the object detection quality. Object detection is performed by applying two-stage data segmentation algorithm based on mean shift clustering. The results quality is measured using structural similarity index and peak signal-to-noise ratio.

User-dependent gesture recognition on android handheld devices

Nowadays, many mobile devices are equipped with built-in inertial sensors. This spurred the research on new forms of communication between man and machines based on the movements or “gestures” performed by the user when holding the device. Here we discuss a gesture recognition system for controlling mobile devices with a wide range of possible practical applications. The system was designed to run in realtime on a resource-constrained platform and therefore has a low computational complexity. The paper describes a GestWiz user application for Android operating system which uses the data from a single triaxial accelerometer to recognize a collection of 9 different hand gestures. The systems performance was evaluated off-line, using a gesture dataset, and on-line, through the series of user tests with the application being executed on a smartphone.

Robust Nonlinear Regression in Enzyme Kinetic Parameters Estimation

Accurate estimation of essential enzyme kinetic parameters, such as and , is very important in modern biology. To this date, linearization of kinetic equations is still widely established practice for determining these parameters in chemical and enzyme catalysis. Although simplicity of linear optimization is alluring, these methods have certain pitfalls due to which they more often then not result in misleading estimation of enzyme parameters. In order to obtain more accurate predictions of parameter values, the use of nonlinear least-squares fitting techniques is recommended. However, when there are outliers present in the data, these techniques become unreliable. This paper proposes the use of a robust nonlinear regression estimator based on modified Tukey’s biweight function that can provide more resilient results in the presence of outliers and/or influential observations. Real and synthetic kinetic data have been used to test our approach. Monte Carlo simulations are performed to illustrate the efficacy and the robustness of the biweight estimator in comparison with the standard linearization methods and the ordinary least-squares nonlinear regression. We then apply this method to experimental data for the tyrosinase enzyme (EC 1.14.18.1) extracted from Solanum tuberosum, Agaricus bisporus, and Pleurotus ostreatus. The results on both artificial and experimental data clearly show that the proposed robust estimator can be successfully employed to determine accurate values of and .

Gradient Compressive Sensing for Image Data Reduction in UAV Based Search and Rescue in the Wild

Search and rescue operations usually require significant resources, personnel, equipment, and time. In order to optimize the resources and expenses and to increase the efficiency of operations, the use of unmanned aerial vehicles (UAVs) and aerial photography is considered for fast reconnaissance of large and unreachable terrains. The images are then transmitted to control center for automatic processing and pattern recognition. Furthermore, due to the limited transmission capacities and significant battery consumption for recording high resolution images, in this paper we consider the use of smart acquisition strategy with decreased amount of image pixels following the compressive sensing paradigm. The images are completely reconstructed in the control center prior to the application of image processing for suspicious objects detection. The efficiency of this combined approach depends on the amount of acquired data and also on the complexity of the scenery observed. The proposed approach is tested on various high resolution aerial images, while the achieved results are analyzed using different quality metrics and validation tests. Additionally, a user study is performed on the original images to provide the baseline object detection performance.

LMNN metric learning and fuzzy nearest neighbour classifier for hand gesture recognition

This paper presents a novel gesture recognition system using a single three-axis accelerometer, that is to serve as an alternative or supplementary interaction modality for controlling mobile devices. Capturing, training and classification of the detected hand gestures are expected to be executed in their entirety on the mobile device running the proposed system, instead of being passed to a nearby computer. As gesture recognition belongs to the group of pattern recognition problems where the underlying class probabilities are not a priori known, the classification is based on the distance between neighbouring examples. The distance metric is optimized by using large margin nearest neighbour (LMNN) method. To measure the amount of classification confidence, a fuzzy version of nearest neighbour algorithm is employed. Obtained results for recognition of nine hand gestures using proposed LMNN—fuzzy combination are presented and compared to that of other similar approaches. The system achieves near perfect recognition accuracy that is highly competitive with systems based on statistical methods and other accelerometer-based gesture recognition systems in the literature.

Motion-Based Gesture Recognition Algorithms for Robot Manipulation

The prevailing trend of integrating inertial sensors in consumer electronics devices has inspired research on new forms of human-computer interaction utilizing hand gestures, which may be set-up on mobile devices themselves. At present, motion gesture recognition is intensely studied, with various recognition techniques being employed and tested. This paper provides an in-depth, unbiased comparison of different algorithms used to recognize gestures based primarily on the single 3D accelerometer recordings. The study takes two of the most popular and arguably the best recognition methods currently in use - dynamic time warping and hidden Markov model - and sets them against a relatively novel approach founded on distance metric learning. The three selected algorithms are evaluated in terms of their overall performance, accuracy, training time, execution time and storage efficacy. The optimal algorithm is further implemented in a prototype user application, aimed to serve as an interface for controlling the motion of a toy robot via gestures made with a smartphone.

A Comparative Study of FFT, DCT, and DWT for Efficient Arrhytmia Classification in RP-RF Framework

Computer-aided ECG classification is an important tool for timely diagnosis of abnormal heart conditions. This paper proposes a novel framework that combines the theory of compressive sensing with random forests to achieve reliable automatic cardiac arrhythmia detection. Furthermore, the paper evaluates the characterization power of FFT, DCT and DWT data transformations in order to extract significant features that will bring the additional boost to the classification performance. The experiments – carried out over MIT-BIH benchmark arrhythmia database, following the standards and recommended practices provided by AAMI – demonstrate that DWT based features exhibit better performances compared to other two feature extraction techniques for a relatively small number of random projected coefficients, i.e. after considerable (approx. 85%) dimensionality reduction of the input signal. The results are very promising, suggesting that the proposed model could be implemented for practical applications of real-time ECG monitoring, due to its low-complexity.

Fair division methods approach as the option of learning process modeling

This paper introduces a new perspective for learning process and student-teacher relationship modeling. A possible framework for including pedagogy in such relationships, which can be applied to both face-to-face and online learning, is nowadays beginning to emerge. This framework combines individual possibilities and resources attributes, confronting them in a fair way. We found game theory - the study of behavioral relations - and fair division methods, their symbols and strategies, useful for modeling and solving our task. Here we suggest and present the original and new idea of modeling learning “contents” as resources, students as players and the teachers as dividers. It is not easy to find a way to fairly share resources between many different players, especially when the resources are continuous. It is not easy to introduce numerical measures to resources-players game (i.e. to create costs). Hence, the divider is confronted with large (theoretically infinite) number of iterative trials looking for appropriate costs and appropriate solution. Therefore the use of computers and software support becomes very important. We have developed first useful application software for this purpose. A simple example of its use is demonstrated in this paper.