Nicola Giulietti

Many neurological conditions, including stroke, Parkinson’s disease, amyotrophic lateral sclerosis, and mild cognitive impairment, compromise cognitive and motor functions and may threaten autonomy in complex daily activities. Rehabilitation can therefore benefit from tasks that simultaneously engage these domains. Driving is a representative example, as it combines continuous motor control, visual processing, attention, decision making, and reaction to unexpected events. Driving simulators can provide a controlled and repeatable environment in which these cognitive and motor functions can be jointly trained, while also supporting the assessment and recovery of driving-related abilities when return to driving is clinically relevant. However, rehabilitation-oriented simulators often provide limited multisensory feedback compared with professional systems, whose size, cost, and complexity limit their transfer to clinical settings. This study presents a compact, three-degree-of-freedom dynamic driving simulator developed for rehabilitation-oriented exergames with adjustable difficulty and measurable in-game metrics. It combines a parallel motion platform, haptic steering, instrumented pedals, and a modular software architecture based on a multibody simulation engine and a graphics engine, with automated session reports. Four exergames were evaluated in 62 healthy licensed adults aged 19–30 years, with each exergame scheduled twice. Traffic Mode was assessed in a separate data-collection campaign involving 38 participants, with partial participant overlap between campaigns. Slalom showed a marked ceiling effect, whereas Linekeeping showed a graded distribution of performance. Reaction Test outcomes varied widely between participants and were strongly associated with approach speed. Participant rank order was moderately preserved across executions. The results characterize the proposed measures in a homogeneous healthy population and identify the most informative metrics and the task configurations requiring refinement. Their suitability for longitudinal monitoring in patients remains to be established.

Accurate measurement of chewing events in natural eating conditions is important for unobtrusive monitoring of feeding behavior and masticatory function. Yet, existing methods often rely on contact sensors, dedicated wearables, or manual annotation. This work presents a non-contact, video-based framework for chewing-event detection using frontal facial video, normalized 3D facial landmark dynamics, and recurrent temporal modeling. To obtain physiologically grounded reference labels, synchronized bilateral anterior temporalis surface electromyography was acquired during real-meal sessions and used to derive chewing-event annotations during dataset construction, whereas inference relied exclusively on video. Facial motion was represented from frame-wise 3D landmarks and processed by recurrent neural networks, with model selection performed through Bayesian hyperparameter optimization. On an independent hold-out test set comprising five sessions and 18,836 frames, the proposed method detected 577 chewing events versus 589 ground truth events, corresponding to a mean absolute error of 4.4 chews/session and a mean absolute percentage error of 4.32%. A comparison with a related rule-based video method from the literature showed substantially larger counting errors (MAE = 39.4, MAPE = 30.39%), particularly in sessions that included concurrent activities such as speaking, suggesting that the proposed approach can reduce counting errors relative to the considered rule-based baseline under the specific meal conditions tested in this feasibility study. The effect of landmark-localization uncertainty on the predicted chewing probability was assessed through Monte Carlo propagation, showing limited impact for most prediction instants and greater sensitivity for intermediate probability values. Finally, the ONNX implementation achieved a mean latency of 8.96 ± 5.74 ms on CPU and 6.89 ± 3.58 ms with CUDA execution on the test workstation, supporting real-time applicability. To support practical deployment, the pipeline was also implemented as a native Kotlin Android application and tested on a commercial tablet, achieving real-time operation at 20 fps.

Accurate and dynamic assessment of human posture during sports activities, such as cycling, can prove useful in optimizing setup, improving performance and reducing the risk of injury. This study presents a metrological evaluation of a marker-less, multi-camera 3D human pose estimation system for cyclist posture analysis, using an optoelectronic motion capture system (OptiTrack) as ground truth. The proposed marker-less system integrates 2D keypoint detections with a Weighted Triangulation method for 3D reconstruction. The experimental protocol involved synchronized acquisition with seven monochrome cameras and OptiTrack reference data during controlled indoor cycling trials. The proposed method is evaluated in terms of Mean Per Joint Position Error, Mean Per Single Joint Position Error, Mean Absolute Joint Angle Error and Pearson Correlation Coefficient. The system achieved an overall average joint position error of 18.3 mm, while joint angle estimation errors ranged from 3.0° to 11.4°, depending on the specific joint and side of the body. The average processing time of 15 ms per frame confirms the suitability of the marker-less system for real-time applications.

Accurate estimation of hazelnut yield is crucial for optimizing resource management and harvest planning. Although the number of female flowers on a flowering plant is a reliable indicator of annual production, counting them remains difficult because of their extremely small size and inconspicuous shape and color. Currently, manual flower counting is the only available method, but it is time-consuming and prone to errors. In this study, a novel vision-based method for automatic flower counting specifically designed for hazelnut plants (Corylus avellana) exploiting a commercial high-resolution imaging system and an image-tiling strategy to enhance small-object detection is proposed. The method is designed to be fast and scalable, requiring less than 8 s per plant for processing, in contrast to 30–60 min typically required for manual counting by human operators. A dataset of 2000 labeled frames was used to train and evaluate multiple female hazelnut flower detection models. To improve the detection of small, low-contrast flowers, a modified YOLO11x architecture was introduced by adding a P2 layer, improving the preservation of fine-grained spatial information and resulting in a precision of 0.98 and a Mean Average Precision (mAP@50-95) of 0.89. The proposed method has been validated on images collected from hazelnut groves and compared with manual counting by four experienced operators in the field, demonstrating its ability to detect small, low-contrast flowers despite occlusions and varying lighting conditions. A regression-based bias correction was applied to compensate for systematic counting deviations, further improving accuracy and reducing the mean absolute percentage error to 27.44%, a value comparable to the variability observed in manual counting. The results indicate that the system can provide a scalable and efficient alternative to traditional female flower manual counting methods, offering an automated solution tailored to the unique challenges of hazelnut yield estimation.

This paper aims to discuss how Thermal Imaging could benefit of Deep-Learning approaches for increasing temperature measurement accuracy in those complex applications where thermal images are subject to severe temperature variations and the target component position varies continuously. Real-time temperature monitoring of braking components during train braking operations indeed represents a well-matched example, given the challenges embedded by this application case: relative camera-to-target movements, variable contrast levels in the thermal image, etc. just to cite some. The advantage of using a specifically trained deep-learning based semantic segmentation model to accurately locate and track target components, is demonstrated by comparing the thermal evolution of a specific component with respect to those obtained exploiting conventional machine vision tracking methods. Experimental validation in a high-speed railway environment shows the proposed method reduces temperature bias to less than 0.5°C when compared to thermocouple data. This accuracy is crucial for ensuring the safety and reliability of railway braking systems. While the widespread use of thermal cameras on trains may be impractical, the approach demonstrates that it is possibile to effectively exploit semantic segmentation models to enhance Thermal Imaging and retrieve more accurate temperature data in complex real-time applications.