Letizia Marchegiani

Scheduled perception for energy-efficient path following

This paper explores the idea of reducing a robots energy consumption while following a trajectory by turning off the main localisation subsystem and switching to a lower-powered, less accurate odometry source at appropriate times. This applies to scenarios where the robot is permitted to deviate from the original trajectory, which allows for energy savings. Sensor scheduling is formulated as a probabilistic belief planning problem. Two algorithms are presented which generate feasible perception schedules: the first is based upon a simple heuristic; the second leverages dynamic programming to obtain optimal plans. Both simulations and real-world experiments on a planetary rover prototype demonstrate over 50% savings in perception-related energy, which translates into a 12% reduction in total energy consumption.

How efficient is estimation with missing data

In this paper, we present a new evaluation approach for missing data techniques (MDTs) where the efficiency of those are investigated using listwise deletion method as reference. We experiment on classification problems and calculate misclassification rates (MR) for different missing data percentages (MDP) using a missing completely at random (MCAR) scheme. We compare three MDTs: pairwise deletion (PW), mean imputation (MI) and a maximum likelihood method that we call complete expectation maximization (CEM). We use a synthetic dataset, the Iris dataset and the Pima Indians Diabetes dataset. We train a Gaussian mixture model (GMM). We test the trained GMM for two cases, in which test dataset is missing or complete. The results show that CEM is the most efficient method in both cases while MI is the worst performer of the three. PW and CEM proves to be more stable, in particular for higher MDP values than MI.

What to measure next to improve decision making? On top-down task driven feature saliency

Top-down attention is modeled as decision making based on incomplete information. We consider decisions made in a sequential measurement situation where initially only an incomplete input feature vector is available, however, where we are given the possibility to acquire additional input values among the missing features. The procecure thus poses the question what to do next? We take an information theoretical approach implemented for generality in a generative mixture model. The framework allows us reduce the decision about what to measure next in a classification problem to the estimation of a few one-dimensional integrals per missing feature. We demonstrate the viability of the framework on four well-known classification problems.

Privacy Leakage of Physical Activity Levels in Wireless Embedded Wearable Systems

With the ubiquity of sensing technologies in our personal spaces, the protection of our privacy and the confidentiality of sensitive data becomes a major concern. In this letter, we focus on wearable embedded systems that communicate data periodically over the wireless medium. In this context, we demonstrate that private information about the physical activity levels of the wearer can leak to an eavesdropper through the physical layer. Indeed, we show that the physical activity levels strongly correlate with changes in the wireless channel that can be captured by measuring the signal strength of the eavesdropped frames. We practically validate this correlation in several scenarios in a real residential environment, using data collected by our prototype wearable accelerometer-based sensor. Finally, we propose a privacy enhancement algorithm that mitigates the leakage of this private information.

The Role of Top-Down Attention in the Cocktail Party: Revisiting Cherry's Experiment after Sixty Years

We investigate the role of top-down task drive attention in the cocktail party problem. In a recently proposed computational model of top-down attention it is possible to simulate the cocktail party problem and make predictions about sensitivity to confounders under different levels of attention. Based on such simulations we expect that under strong top-down attention pattern recognition is improved as the model can compensate for noise and confounders. We next investigate the role of temporal and spectral overlaps and speech intelligibility in humans, and how the presence of a task influences their relation. For this purpose, we perform behavioral experiments inspired by Cherrys classic experiments carried out almost sixty years ago. We make participants listen to a mono signal consisting of two different narratives pronounced by a speech synthesizer under two different conditions. In the first case, participants listen with no specific task, while in the second one they are asked to follow one of the stories. Participants report the words they heard by choosing from a list which also includes terms not present in any of the narratives. We define temporal and spectral overlaps using the ideal binary mask (IBMs) as a gauge. We analyze the correlation between overlaps and the amount of reported words. We observe a significant negative correlation when there is no task, while no correlation is detected when a task is involved. Hence, results that are well aligned with the simulation results in our computational top-down attention model.

Leveraging the urban soundscape: Auditory perception for smart vehicles

Urban environments are characterised by the presence of distinctive audio signals which alert the drivers to events that require prompt action. The detection and interpretation of these signals would be highly beneficial for smart vehicle systems, as it would provide them with complementary information to navigate safely in the environment. In this paper, we present a framework that spots the presence of acoustic events, such as horns and sirens, using a two-stage approach. We first model the urban soundscape and use anomaly detection to identify the presence of an anomalous sound, and later determine the nature of this sound. As the audio samples are affected by copious non-stationary and unstructured noise, which can degrade classification performance, we propose a noise-removal technique to obtain a clean representation of the data we can use for classification and waveform reconstruction. The method is based on the idea of analysing the spectrograms of the incoming signals as images and applying spectrogram segmentation to isolate and extract the alerting signals from the background noise. We evaluate our framework on four hours of urban sounds collected driving around urban Oxford on different kinds of road and in different traffic conditions. When compared to traditional feature representations, such as Mel-frequency cepstrum coefficients, our framework shows an improvement of up to 31% in the classification rate.

On cross-language consonant identification in second language noise

Speech perception in everyday conditions is highly affected by the presence of noise of a different nature. The presence of overlapping speakers is considered an especially challenging scenario, as it introduces both energetic and informational masking. The efficacy of the masking also depends on the familiarity with the language of both the target and masking stimuli. This work analyses consonant identification by non-native English speakers in N-talker natural babble noise and babble-modulated noise, by varying the number of talkers in the babble. In particular, only English consonants that are also present in all the native languages of the subjects are used. As the subjects are familiar with the consonants used, this study can be considered a step towards a deeper analysis on perception of first language speech in the presence of second language maskers.

Top-down attentionwith features missing at random

In this paper we present a top-down attention model designed for an environment in which features are missing completely at random. Following (Hansen et al., 2011) we model top-down attention as a sequential decision making process driven by a task - modeled as a classification problem - in an environment with random subsets of features missing, but where we have the possibility to gather additional features among the ones that are missing. Thus, the top-down attention problem is reduced to finding the answer to the question what to measure next? Attention is based on the top-down saliency of the missing features given as the estimated difference in classification confusion (entropy) with and without the given feature. The difference in confusion is computed conditioned on the available set of features. In this work, we make our attention model more realistic by also allowing the initial training phase to take place with incomplete data. Thus, we expand the model to include a missing data technique in the learning process. The top-down attention mechanism is implemented in a Gaussian Discrete mixture model setting where marginals and conditionals are relatively easy to compute. To illustrate the viability of expanded model, we train the mixture model with two different datasets, a synthetic data set and the well-known Yeast dataset of the UCI database. We evaluate the new algorithm in environments characterized by different amounts of incompleteness and compare the performance with a system that decides next feature to be measured at random. The proposed top-down mechanism clearly outperforms random choice of the next feature.

Learning to Listen to Your Ego-(motion): Metric Motion Estimation from Auditory Signals.

This paper is about robot ego-motion estimation relying solely on acoustic sensing. By equipping a robot with microphones, we investigate the possibility of employing the noise generated by the motors and actuators of the vehicle to estimate its motion. Audio-based odometry is not affected by the scene’s appearance, lighting conditions, and structure. This makes sound a compelling auxiliary source of information for ego-motion modelling in environments where more traditional methods, such as those based on visual or laser odometry, are particularly challenged. By leveraging multi-task learning and deep architectures, we provide a regression framework able to estimate the linear and the angular velocity at which the robot has been travelling. Our experimental evaluation conducted on approximately two hours of data collected with an unmanned outdoor field robot demonstrated an absolute error lower than 0.07 m/s and 0.02 rad/s for the linear and angular velocity, respectively. When compared to a baseline approach, making use of single-task learning scheme, our system shows an improvement of up to 26% in the ego-motion estimation.

Enabling intelligent energy management for robots using publicly available maps

Energy consumption represents one of the most basic constraints for mobile robot autonomy. We propose a new framework to predict energy consumption using information extracted from publicly available maps. This method avoids having to model internal robot configurations, which are often unavailable, while still providing invaluable predictions for both explored and unexplored trajectories. Our approach uses a heteroscedastic Gaussian Process to model the power consumption, which explicitly accounts for variance due to exogenous latent factors such as traffic and weather conditions. We evaluate our framework on 30km of data collected from a city centre environment with a mobile robot travelling on pedestrian walkways. Results across five different test routes show an average difference between predicted and measured power consumption of 3.3%, leading to an average error of 6.6% on predictions of energy consumption. The distinct advantage of our model is our ability to predict measurement variance. The variance predictions improved by 84.3% over a benchmark.