Autonomous Driving (AD) entails vehicles that can perceive their surroundings and navigate without human intervention. This involves utilising a combination of sensors and algorithms to recognise obstacles, interpret traffic signals, and make driving decisions. While AD holds promise for transforming transportation by enhancing safety, reducing congestion, minimising pollution, and optimising efficiency, it poses technical challenges also. This work extends a novel approach to building an autonomous vehicle agent using Deep Reinforcement Learning (DRL) with Proximal Policy Optimisation (PPO) to navigate urban environments simulated by the CAR Learning to Act (CARLA) Simulator. The agent aims to maintain lane integrity and avoid collisions, even in adverse weather conditions. The proposed architecture integrates a 180-degree environmental view and various multimodal data inputs (RGB, segmentation, and depth camera inputs), extensively tested through experimentation. Notably, the integration of segmentation and depth data results in a 13% reduction in the collision rate, with the proposed agent achieving a total reward of 2510. This approach demonstrates significant progress over the previous framework, showcasing improved obstacle detection and collision avoidance accuracy. Moreover, these findings contribute to ongoing autonomous vehicle research, offering insights into effective strategies for developing robust and dependable driving agents capable of navigating urban environments and interacting with road infrastructure, contributing to advancements in Augmented Intelligence of Things (AIoT)-enabled autonomous driving.

Abstract Ultrasound is a readily available, non-invasive and low-cost screening for the identification of endometriosis lesions, but its diagnostic specificity strongly depends on the experience of the operator. For this reason, computer-aided diagnosis tools based on Artificial Intelligence techniques can provide significant help to the clinical staff, both in terms of workload reduction and in increasing the overall accuracy of this type of examination and its outcome. However, although these techniques are spreading rapidly in a variety of domains, their application to endometriosis is still very limited. To fill this gap, we propose and evaluate a novel multi-scale ensemble approach for the automatic segmentation of endometriosis lesions from transvaginal ultrasounds. The peculiarity of the method lies in its high discrimination capability, obtained by combining, in a fusion fashion, multiple Convolutional Neural Networks trained on data at different granularity. The experimental validation carried out shows that: (i) the proposed method allows to significantly improve the performance of the individual neural networks, even in the presence of a limited training set; (ii) with a Dice coefficient of 82%, it represents a valid solution to increase the diagnostic efficacy of the ultrasound examination against such a pathology.

Abstract In the last years, scientific and industrial research has experienced a growing interest in acquiring large annotated data sets to train artificial intelligence algorithms for tackling problems in different domains. In this context, we have observed that even the market for football data has substantially grown. The analysis of football matches relies on the annotation of both individual players’ and team actions, as well as the athletic performance of players. Consequently, annotating football events at a fine-grained level is a very expensive and error-prone task. Most existing semi-automatic tools for football match annotation rely on cameras and computer vision. However, those tools fall short in capturing team dynamics and in extracting data of players who are not visible in the camera frame. To address these issues, in this manuscript we present FootApp, an AI-based system for football match annotation. First, our system relies on an advanced and mixed user interface that exploits both vocal and touch interaction. Second, the motor performance of players is captured and processed by applying machine learning algorithms to data collected from inertial sensors worn by players. Artificial intelligence techniques are then used to check the consistency of generated labels, including those regarding the physical activity of players, to automatically recognize annotation errors. Notably, we implemented a full prototype of the proposed system, performing experiments to show its effectiveness in a real-world adoption scenario.

Abstract Emotions recognition is widely applied for many tasks in different fields, from human-computer and human-robot interaction to learning platforms. Also, it can be used as an intrinsic approach for face recognition tasks, in which an expression-independent face classifier is developed. Most approaches face the problem by designing deeper and deeper neural networks that consider an expression as a still image or, in some cases, a sequence of consecutive frames depicting the temporal component of the expression. However, these suffer the training phase’s computational burden, which can take hours or days to be completed. In this work, a Web Shaped Model is proposed, which consists of a geometrical approach for extracting discriminant features from a face, depicting the characteristics of an expression. The model does not need to be trained since it is applied on a face and centred on the nose tip, resulting in image size and face size independence. Experiments on publicly available datasets show that this approach reaches comparable and even better results than those obtained applying DNN-based approaches.

In this work, we combine Curriculum Learning with Deep Reinforcement Learning to learn without any prior domain knowledge, an end-to-end competitive driving policy for the CARLA autonomous driving simulator. To our knowledge, we are the first to provide consistent results of our driving policy on all towns available in CARLA. Our approach divides the reinforcement learning phase into multiple stages of increasing difficulty, such that our agent is guided towards learning an increasingly better driving policy. The agent architecture comprises various neural networks that complements the main convolutional backbone, represented by a ShuffleNet V2. Further contributions are given by (i) the proposal of a novel value decomposition scheme for learning the value function in a stable way and (ii) an ad-hoc function for normalizing the growth in size of the gradients. We show both quantitative and qualitative results of the learned driving policy.

Autonomous Vehicles promise to transport people in a safer, accessible, and even efficient way. Nowadays, real-world autonomous vehicles are build by large teams from big companies with a tremendous amount of engineering effort. Deep Reinforcement Learning can be used instead, without domain experts, to learn end-to-end driving policies. Here, we combine Curriculum Learning with deep reinforcement learning, in order to learn without any prior domain knowledge, an end-to-end competitive driving policy for the CARLA autonomous driving simulator. To our knowledge, this is the first work which provides consistent results of our driving policy on all the town scenarios provided by CARLA. Moreover, we point out two important issues in reinforcement learning: the former is about learning the value function in a stable way, whereas the latter is related to normalizing the learned advantage function. A proposal of a solution to these problems is provided.

Manual event tagging may be a very long and stressful activity, due the monotonous operations involved. This is particularly true when dealing with online video tagging, as for football matches, in which the burden of events to tag can consist of many thousands of actions, according to the desired level of granularity. In this work we describe an actual solution, developed for an existing football match tagging application, in which the GUI has been enhanced and integrated with a Voice User Interface, aiming at reducing tagging time and error rate. Empirical tests have revealed the efficiency and the benefits brought by the developed solution.

Head pose estimation is a sensitive topic in video surveillance/smart ambient scenarios since head rotations can hide/distort discriminative features of the face. Face recognition would often tackle the problem of video frames where subjects appear in poses making it quite impossible. In this respect, the selection of the frames with the best face orientation can allow triggering recognition only on these, therefore decreasing the possibility of errors. This paper proposes a novel approach to head pose estimation for smart cities and video surveillance scenarios, aiming at this goal. The method relies on a cascade of two models: the first one predicts the positions of 68 well-known face landmarks; the second one applies a web-shaped model over the detected landmarks, to associate each of them to a specific face sector. The method can work on detected faces at a reasonable distance and with a resolution that is supported by several present devices. Results of experiments executed over some classical pose estimation benchmarks, namely Point '04, Biwi, and AFLW datasets show good performance in terms of both pose estimation and computing time. Further results refer to noisy images that are typical of the addressed settings. Finally, examples demonstrate the selection of the best frames from videos captured in video surveillance conditions.

The analysis of the shape and geometry of the human hand has long represented an attractive field of research to address the needs of digital image forensics. Over recent years, it has also turned out to be effective in biometrics, where several innovative research lines are pursued. Given the widespread diffusion of mobile and portable devices, the possibility of checking the owner identity and controlling the access to the device by the hand image looks particularly attractive and less intrusive than other biometric traits. This encourages new research to tacklethe present limitations. The proposed work implements the complete architecture of a mobile hand recognition system, which uses the camera of the mobile device for the acquisition of the hand in visible light spectrum. The segmentation of the hand starts from the detection of the convexities and concavities defined by the fingers, and allows extracting 57 different features from the hand shape. The main contributions of the paper develop along two directions. First, dimensionality reduction methods are investigated, in order to identify subsets of features including only the most discriminating and robust ones. Second, different matching strategies are compared. The proposed method is tested over a dataset of hands from 100 subjects. The best obtained Equal Error Rate is 0.52%. Results demonstrate that discarding features that are more prone to distortions allows lighter processing, but also produces better performance than using the full set of features. This confirms the feasibility of such an approach on mobile devices, and further suggests to adopt it even in more traditional settings.

Context awareness is major component of Ambient Intelligence. In fact, Ambient Intelligent environments are designed to combine ubiquity, awareness, intelligence and natural interaction. Awareness is defined as the ability by the system to locate and recognize people and objects, and their intentions. Then, intelligence is the ability of the system to analyze the detected context and to adapt its behavior to people and situations, and to learn over time, in order to provide users with personalized services. These concepts date back to late'90, but nowadays the widespread and ubiquitous availability of mobile devices, equipped with several different sensors, allows to put them into practice in a number of unexpected ways. This work presents a preliminary investigation on the possibility to use some of the smartphone sensors, namely the accelerometer and the gyroscope, to identify the bodily context when the user lifts the device to answer a call. The arm gesture, i.e., the way it is performed, is classified into 4 different states: while standing, sitting, walking or running. This information can be used to trigger context-sensitive system actions.

In the last fifteen years, smartphones have become very popular amongst the population, with the subsequent development of dozens of applications aimed at providing security to these portable devices. Nowadays, the cutting edge devices are also provided with biometric sensors (e.g., fingerprint sensors) allowing the users to access them without using the out-of-date alphanumerical password. In this work, we present a method that realizes iris recognition by means of Self Organizing Maps (SOM). In order to obtain a better refined and discriminative feature map, the RGB data of the iris, previously segmented, have been combined with two statistical descriptors. The algorithm has been designed specifically to require a low processing power, making it an ideal choice in the context of mobile devices.

Soft biometrics have been emerging to complement other traits and are particularly useful for poor quality data. In this paper, we propose an efficient algorithm to estimate human head poses and to infer soft biometric labels based on the 3D morphology of the human head. Starting by considering a set of pose hypotheses, we use a learning set of head shapes synthesized from anthropometric surveys to derive a set of 3D head centroids that constitutes a metric space. Next, representing queries by sets of 2D head landmarks, we use projective geometry techniques to rank efficiently the joint 3D head centroids/pose hypotheses according to their likelihood of matching each query. The rationale is that the most likely hypotheses are sufficiently close to the query, so a good solution can be found by convex energy minimization techniques. Once a solution has been found, the 3D head centroid and the query are assumed to have similar morphology, yielding the soft label. Our experiments point toward the usefulness of the proposed solution, which can improve the effectiveness of face recognizers and can also be used as a privacy-preserving solution for biometric recognition in public environments.

Multi-biometric systems are a powerful solution to deal with limitations of single classifiers, therefore improving the final recognition accuracy. The sub-systems composing the final architecture often return supplementary indices of input quality and/or of response reliability, which further qualify each recognition score. These indices can enter different information fusion policies. First, they can be used as weights for the fusion of the corresponding scores, in such a way that less trustworthy responses have a lower influence. Alternatively, they can be used to drive the selection of a subset of systems actually enabled for each fusion operation. The present work discusses their appropriate combination with respective scores, to obtain single values which are easier to handle and compare. It is worth underlining the different nature of quality and reliability measures. The quality estimation of input samples requires a complex analysis of environmental conditions, including capture sensors, besides computations over acquired data. Reliability of a system estimates its ability to return a correct response. As an alternative to combination, some solutions rather estimate the joint distributions of conditional probabilities of the scores from the single subsystems. These solutions require training through a huge number of samples. Furthermore, they assume stable score distributions. Our unified representation of the recognition score and of the corresponding quality/reliability value into a single complex number provides simplification and speed up of fusion of multi-classifier results. It also allows to devise procedures to readily compare the performance of different modules in a multi-biometric system, given that there is no natural ordering of these pairs of values of different nature. Moreover our method achieves performance comparable to top performing schemes, yet does not require a prior estimation of (joint) score distributions. As a matter of fact, though representing an upper bound to the obtainable performance, Likelihood ratio has the limit to require an accurate estimation of score distributions, while our approach relies on the reliability of each single response. This feature is very interesting when the set of relevant subjects may present significant variations over time.