Abstract

Vision based active safety systems have become more frequently occurring in modern vehicles to estimate depth of the objects ahead and for autonomous driving (AD) and advanced driver-assistance systems (ADAS). In this thesis a lightweight deep convolutional neural network performing real-time depth estimation on single monocular images is implemented and evaluated. Many of the vision based automatic brake systems in modern vehicles only detect pre-trained object types such as pedestrians and vehicles. These systems fail to detect general objects such as road debris and roadside obstacles. In stereo vision systems the problem is resolved by calculating a disparity image from the stereo image pair to extract depth information. The distance to an object can also be determined using radar and LiDAR systems. By using this depth information the system performs necessary actions to avoid collisions with objects that are determined to be too close. However, these systems are also more expensive than a regular mono camera system and are therefore not very common in the average consumer car. By implementing robust depth estimation in mono vision systems the benefits from active safety systems could be utilized by a larger segment of the vehicle fleet. This could drastically reduce human error related traffic accidents and possibly save many lives. The network architecture evaluated in this thesis is more lightweight than other CNN architectures previously used for monocular depth estimation. The proposed architecture is therefore preferable to use on computationally lightweight systems. The network solves a supervised regression problem during the training procedure in order to produce a pixel-wise depth estimation map. The network was trained using a sparse ground truth image with spatially incoherent and discontinuous data and output a dense spatially coherent and continuous depth map prediction. The spatially incoherent ground truth posed a problem of discontinuity that was addressed by a masked loss function with regularization. The network was able to predict a dense depth estimation on the KITTI dataset with close to state-of-the-art performance. 

Abstract

End-to-end speech synthesis is a promising approach that directly converts raw text to speech. Although it was shown that Tacotron2 outperforms classical pipeline systems with regards to naturalness in English, its applicability to other languages is still unknown. Japanese could be one of the most difficult languages for which to achieve end-to-end speech synthesis, largely due to its character diversity and pitch accents. Therefore, state-of-the-art systems are still based on a traditional pipeline framework that requires a separate text analyzer and duration model. Towards end-to-end Japanese speech synthesis, we extend Tacotron to systems with self-attention to capture long-term dependencies related to pitch accents and compare their audio quality with classical pipeline systems under various conditions to show their pros and cons. In a large-scale listening test, we investigated the impacts of the presence of accentual-type labels, the use of force or predicted alignments, and acoustic features used as local condition parameters of the Wavenet vocoder. Our results reveal that although the proposed systems still do not match the quality of a top-line pipeline system for Japanese, we show important stepping stones towards end-to-end Japanese speech synthesis. Comment: to be appeared at ICASSP 2019

Abstract

International audience; In this work, we investigate the use of OpenStreetMap data for semantic labeling of Earth Observation images. Deep neural networks have been used in the past for remote sensing data classification from various sensors, including multispectral, hyperspectral, SAR and LiDAR data. While OpenStreetMap has already been used as ground truth data for training such networks, this abundant data source remains rarely exploited as an input information layer. In this paper, we study different use cases and deep network architectures to leverage OpenStreetMap data for semantic labeling of aerial and satellite images. Especially , we look into fusion based architectures and coarse-to-fine segmentation to include the OpenStreetMap layer into multispectral-based deep fully convolutional networks. We illustrate how these methods can be successfully used on two public datasets: ISPRS Potsdam and DFC2017. We show that OpenStreetMap data can efficiently be integrated into the vision-based deep learning models and that it significantly improves both the accuracy performance and the convergence speed of the networks.

Abstract

This presentation was aimed to explore the growing opportunities to merge two booming fields: deep learning and autonomous vehicles, from a technical point of view. It addressed some Intelligent Systems Laboratory (Universidad Carlos III de Madrid, Spain) developments in this line of research, such as an obstacle detection framework using convolutional neural networks (CNNs). Furthermore, it presented a large number of challenging driving-related tasks that were expected to become tractable through this new approach, with the focus on the strong requirements posed by the upcoming self-driving systems."/p" "p"This presentation was part of the 6th LSI Ph.D. Meeting, which was held on 14 Jun 2016 at the Escuela Politécnica Superior of the Universidad Carlos III de Madrid. It was published on Zenodo as an exercise within the THOR Bootcamp on Open Data, organized on 16 Nov 2016.

Abstract

Autonomous vehicles (AVs) are already operating on the streets of many countries around the globe. Contemporary concerns about AVs do not relate to the implementation of fundamental technologies, as they are already in use, but are rather increasingly centered on the way that such technologies will affect emerging transportation systems, our social environment, and the people living inside it. Many concerns also focus on whether such systems should be fully automated or still be partially controlled by humans. This work aims to address the new reality that is formed in autonomous shuttles mobility infrastructures as a result of the absence of the bus driver and the increased threat from terrorism in European cities. Typically, drivers are trained to handle incidents of passengers’ abnormal behavior, incidents of petty crimes, and other abnormal events, according to standard procedures adopted by the transport operator. Surveillance using camera sensors as well as smart software in the bus will maximize the feeling and the actual level of security. In this paper, an online, end-to-end solution is introduced based on deep learning techniques for the timely, accurate, robust, and automatic detection of various petty crime types. The proposed system can identify abnormal passenger behavior such as vandalism and accidents but can also enhance passenger security via petty crimes detection such as aggression, bag-snatching, and vandalism. The solution achieves excellent results across different use cases and environmental conditions. Document type: Article

Abstract

Environmental sensors are exploited in smart homes for many purposes. Sensor data inherently carries behavioral information, possibly useful to infer wellness and health-related insights in an indirect fashion. In order to exploit such features, however, powerful analytics are needed to convert raw sensor output into meaningful and accessible knowledge. In this paper, a complete monitoring architecture is presented, including home sensors and cloud-based back-end services. Unsupervised techniques for behavioral data analysis are presented, including: (i) regression and outlier detection models (also used as feature extractors for more complex models) (ii) statistical hypothesis testing frameworks for detecting changes in sensor-detected activities and (iii) a clustering process, leveraging deep learning techniques, for extracting complex, multivariate patterns from daily sensor data. Such methods are discussed and evaluated on real-life data, collected within several EU-funded projects. Overall, the presented methods may prove very useful to build effective monitoring services, suitable for practical exploitation in caregiving activities, complementing conventional telemedicine techniques. Document type: Article

Abstract

In order to improve the efficiency of transportation networks, it is critical to forecast traffic congestion. Large-scale traffic congestion data have become available and accessible, yet they need to be properly represented in order to avoid overfitting, reduce the requirements of computational resources, and be utilized effectively by various methodologies and models. Inspired by pooling operations in deep learning, we propose a representation framework for traffic congestion data in urban road traffic networks. This framework consists of grid-based partition of urban road traffic networks and a pooling operation to reduce multiple values into an aggregated one. We also propose using a pooling operation to calculate the maximum value in each grid (MAV). Raw snapshots of traffic congestion maps are transformed and represented as a series of matrices which are used as inputs to a spatiotemporal congestion prediction network (STCN) to evaluate the effectiveness of representation when predicting traffic congestion. STCN combines convolutional neural networks (CNNs) and long short-term memory neural network (LSTMs) for their spatiotemporal capability. CNNs can extract spatial features and dependencies of traffic congestion between roads, and LSTMs can learn their temporal evolution patterns and correlations. An empirical experiment on an urban road traffic network shows that when incorporated into our proposed representation framework, MAV outperforms other pooling operations in the effectiveness of the representation of traffic congestion data for traffic congestion prediction, and that the framework is cost-efficient in terms of computational resources. Document type: Article

Abstract

An ultrasonic sensors system is commonly used to measure the wall thickness of buried pipelines in the transportation of oil and gas. The key of the system is to precisely measure time-of-flight difference (TOFD) produced by the reflection of ultrasonic on the inner and outer surfaces of the pipelines. In this paper, based on deep learning, a novel method termed Wave-Transform Network is proposed to tackle the issues. The network consists of two parts: part 1 is designed to separate the potential overlapping ultrasonic echo signals generated from two surfaces, and part 2 is utilized to divide the sample points of each signal into two types corresponding to before and after the arrival time of ultrasonic echo, which can determine the time-of-flight (TOF) of each signal and calculate the thickness of pipelines. Numerical simulation and actual experiments are carried out, and the results show satisfactory performances. Document type: Article

Abstract

Short-term load forecasting is a key task to maintain the stable and effective operation of power systems, providing reasonable future load curve feeding to the unit commitment and economic load dispatch. In recent years, the boost of internal combustion engine (ICE) based vehicles leads to the fossil fuel shortage and environmental pollution, bringing significant contributions to the greenhouse gas emissions. One of the effective ways to solve problems is to use electric vehicles (EVs) to replace the ICE based vehicles. However, the mass rollout of EVs may cause severe problems to the power system due to the huge charging power and stochastic charging behaviors of the EVs drivers. The accurate model of EV charging load forecasting is, therefore, an emerging topic. In this paper, four featured deep learning approaches are employed and compared in forecasting the EVs charging load from the charging station perspective. Numerical results show that the gated recurrent units (GRU) model obtains the best performance on the hourly based historical data charging scenarios, and it, therefore, provides a useful tool of higher accuracy in terms of the hourly based short-term EVs load forecasting. Document type: Article