Jaakko Suutala

Methods for person identification on a pressure-sensitive floor: Experiments with multiple classifiers and reject option

This paper presents methods for footstep-based person identification using a large pressure-sensitive floor with a sensory system. The aim was to analyse and compare different pattern classification methods for their ability to solve this particular problem as well as to introduce some novel and useful methodological extensions, which can improve classification accuracy and the adaptability of the system. These extensions are based on the conditional posterior probability outputs of classifiers, i.e., efforts to combine classifiers trained with different feature sets and to combine multiple footstep instances of a single person walking on the floor. Additionally, a method to reject unreliable examples in order to increase accuracy was applied to the system. The experiments demonstrated the usefulness of these methods. An identification method that uses a combination of multiple classifiers and multiple examples yielded very promising results with an overall accuracy rate of 92% for ten different walkers. When the reject option was added, a classification rate of 95% with a 9% rejection rate was achieved. This methodology can be applied to smart room applications where a small number of persons need to be identified.

Spectrum Occupancy Measurements: A Survey and Use of Interference Maps

In order to provide meaningful data about spectrum use, occupancy measurements describing the utilization rate of a specific frequency band should be conducted over a specific area instead of a single location. This paper presents a comprehensive methodology for the measurement and analysis of spectrum occupancy. This paper surveys spectrum measurement campaigns and associated interference maps, introducing the latter as a tool for spectrum analysis and management based on measurement data. An interference map characterizes the spectrum use by defining the level of interference over an area of interest in a certain frequency band. Building on findings from practical measurement studies, guidelines for spectrum occupancy measurements are given. While many scientific spectrum occupancy measurement papers tend to be too optimistic about the significance and generality of the results, we propose a cautionary perspective on drawing strong conclusions based on the often limited amount of data gathered. The different phases of the spectrum occupancy measurement and analysis process are described and a thorough discussion of interpolation methods is provided. Means to improve the measurement accuracy are discussed, especially regarding spatial domain considerations and the impact of the sampling interval on the results. A practical example of an improved measurement system design covering all the phases of the measurement process and used at the Turku, Finland; Blacksburg, VA, USA; and Chicago, IL, USA, spectrum observatories is given. Using the improved design, more realistic spectrum occupancy data can be obtained to lay the foundation for spectrum management decisions.

Discriminative temporal smoothing for activity recognition from wearable sensors

This paper describes daily life activity recognition using wearable acceleration sensors attached to four different parts of the human body. The experimental data set consisted of signals recorded from 13 different subjects performing 17 daily activities. Furthermore, to attain more general activities, some of the most specific classes were combined for a total of 9 different activities. Simple time domain features were calculated from each sensor device. For the recognition task, we propose a novel sequential learning method that combines discriminative learning of individual input-output mappings using support vector machines (SVM) with generative learning to smooth temporal time-dependent activity sequences with a trained hidden Markov model (HMM) type transition probability matrix. The experiments show that the accuracy of the proposed method is superior to various conventional discriminative and generative methods alone, and it achieved a total recognition rate of 94% and 96% studying 17 and 9 different daily activities, respectively.

Combining classifiers with different footstep feature sets and multiple samples for person identification

Combination of classifiers is usually a good strategy to improve accuracy in pattern recognition systems. In this paper, we present a new approach to footstep-based biometric identification by combining pattern classifiers with different feature sets. Footstep profiles are obtained from a pressure-sensitive floor. Our identification system consists of two different combination stages. At the first stage, three pattern classifiers, trained with feature sets presenting different characteristics of input signal, are combined. The feature sets include the spatial domain properties of the footstep profile as well as the frequency domain presentation of the signal and its derivative. At the second stage, multiple input samples are combined, using the posterior probability outputs from the first stage, to make the final decision. The building blocks of the classification system are examined, and the methodological justifications are analyzed. The experimental results show improvements in identification accuracies compared to previously reported work.

Gaussian Process Person Identifier Based on Simple Floor Sensors

This paper describes methods and sensor technology used to identify persons from their walking characteristics. We use an array of simple binary switch floor sensors to detect footsteps. Feature analysis and recognition are performed with a fully discriminative Bayesian approach using a Gaussian Process (GP) classifier. We show the usefulness of our probabilistic approach on a large data set consisting of walking sequences of nine different subjects. In addition, we extract novel features and analyse practical issues such as the use of different shoes and walking speeds, which are usually missed in this kind of experiment. Using simple binary sensors and the large nine-person data set, we were able to achieve promising identification results: a 64% total recognition rate for single footstep profiles and an 84% total success rate using longer walking sequences (including 5 - 7-footstep profiles). Finally, we present a context-aware prototype application. It uses person identification and footstep location information to provide reminders to a user.

Reject-Optional LVQ-Based Two-Level Classifier to Improve Reliability in Footstep Identification

This paper reports experiments of recognizing walkers based on measurements with a pressure-sensitive EMFi-floor. Identification is based on a two-level classifier system. The first level performs Learning Vector Quantization (LVQ) with a reject option to identify or to reject a single footstep. The second level classifies or rejects a sequence of three consecutive identified footsteps based on the knowledge from the first level. The system was able to reduce classification error compared to a single footstep classifier without a reject option. The results show a 90% overall success rate with a 20% rejection rate, identifying eleven walkers, which can be considered very reliable.

SE-155 DBSA: a device-based software architecture for data mining

In this paper a new architecture for a variety of data mining tasks is introduced. The Device-Based Software Architecture (DBSA) is a highly portable and generic data mining software framework where processing tasks are modeled as components linked together to form a data mining application. The name of the architecture comes from the analogy that each processing task in the framework can be thought of as a device. The framework handles all the devices in the same manner, regardless of whether they have a counterpart in the real world or whether they are just logical devices inside the framework. The DBSA offers many reusable devices, ready to be included in applications, and the application programmer can easily code new devices for the architecture. The framework is bundled with connections to several widely used external tools and languages, making prototyping new applications easy and fast. In the paper we compare DBSA to existing data mining frameworks, review its design and present a case study application implemented with the framework. The paper shows that the DBSA can act as a base for diverse data mining applications.

Instance Level Classification Confidence Estimation

Often the confidence of a classification prediction can be as important as the prediction itself although current classification confidence measures are not necessarily consistent between different data sets. Thus in this paper, we present an algorithm to predict instance level classification confidence that is more consistent between data sets and is intuitive to interpret. The results with five test cases show high correlation between true and predicted classification rate, i.e. the probability of assigning the correct class label, thus proving the validity of the proposed algorithm.

Classification Uncertainty of Multiple Imputed Data

Every classification model contains uncertainty. This uncertainty can be distributed evenly or into certain areas of feature space. In regular classification tasks, the uncertainty can be estimated from posterior probabilities. On the other hand, if the data set contains missing values, not all classifiers can be used directly. Imputing missing values solves this problem but it suppresses variation in the data leading to underestimation of uncertainty and can also bias the results. Multiple imputation, where several copies of the data set are created, solves these problems but the classical approach for uncertainty estimation does not generalize to this case. Thus in this paper we propose a novel algorithm to estimate classification uncertainty with multiple imputed data. We show that the algorithm performs as well as the benchmark algorithm with a classifier that supports classification with missing values. It also supports the use of any classifier, even if it does not support classification with missing values, as long as it supports the estimation of posterior probabilities.

Getting More Out of Small Data Sets - Improving the Calibration Performance of Isotonic Regression by Generating More Data.

Often it is necessary to have an accurate estimate of the probability that a classifier prediction is indeed correct. Many classifiers output a prediction score that can be used as an estimate of that probability but for many classifiers these prediction scores are not well calibrated. If enough training data is available, it is possible to post process these scores by learning a mapping from the prediction scores to probabilities. One of the most used calibration algorithms is isotonic regression. This kind of calibration, however, requires a decent amount of training data to not overfit. But many real world data sets do not have excess amount of data that can be set aside for calibration. In this work, we have developed a data generation algorithm to produce more data from a limited sized training data set. We used two variations of this algorithm to generate the calibration data set for isotonic regression calibration and compared the results to the traditional approach of setting aside part of the training data for calibration. Our experimental results suggest that this can be a viable option for smaller data sets if good calibration is essential.