Improved approximated median filter algorithm for real-time computer vision applications

Machine Learning and 3D Reconstruction of Materials Surface for Nondestructive Inspection

During the steel pipeline installation, special attention is paid to the butt weld control performed by fusion welding. The operation of the currently popular automated X-ray and ultrasonic testing complexes is associated with high resource and monetary costs. In this regard, this work is devoted to the development of alternative and cost-effective means of preliminary quality control of the work performed based on the visual testing method. To achieve this goal, a hardware platform based on a single board Raspberry Pi4 minicomputer and a set of available modules and expansion cards is proposed, and software whose main functionality is implemented based on the systemic application of computer vision algorithms and machine learning methods. The YOLOv5 object detection algorithm and the random forest machine learning model were used as a defect detection and classification system. The mean average precision (mAP) of the trained YOLOv5 algorithm based on extracted weld contours is 86.9%. A copy of YOLOv5 trained on the images of control objects showed a mAP result of 96.8%. Random forest identifying of the defect precursor based on the point clouds of the weld surface achieved a mAP of 87.5%.

Detecting Proximity with Bluetooth Low Energy Beacons for Cultural Heritage

The RE.S.I.STO project targets visitors of Pisa medieval city, with the goal of providing high-quality digital contents accessible with smart devices. We describe the design, implementation and the test phases of the RE.S.I.STO application, whose goal is to automatically detect the proximity between visitors and artworks. Proximity is detected with a set of algorithms based on the analysis of Bluetooth Low Energy beacons. We detail our experimental campaigns which reproduce several museum layouts of increasing complexity at two pilot sites, and we compute the performance of the implemented algorithms to detect the nearby artworks. In particular, we test our solution in a wide open space located in our research institute and by performing a real deployment at the Camposanto Monumentale located in Pisa (Italy). The obtained performance varies in the range of 40% to perfect accuracy, according to the complexity of the considered museum layouts. We also describe a set of stress and stability tests aimed at verifying the robustness of the application during the data collection process. Our results show that the mobile application is able to reduce the beacon loss rate, with an average value of 77% of collected beacons.