Design and Development of Real-Time Assistive Algorithms for Individuals with Sensory Impairments

The aim of this thesis is to develop innovative algorithms to support individuals with auditory impairments, enhancing their autonomy and safety through intelligent embedded systems. Specifically, the system is designed to detect and classify sounds associated with potentially dangerous events, providing timely visual notifications that alert users in real-world situations. These algorithms process auditory input to enrich sensory information and deliver real-time assistance through intuitive feedback mechanisms. The implementation is based on the STMicroelectronics STM32U5-based B-U585I-IOT02A board and the Corellium virtualization environment, which together enable accurate simulation, testing and optimization before deployment. Particular attention has been given to optimizing execution time and memory usage, ensuring real-time hazard detection and efficient resource management on a low-power microcontroller platform. A further goal of this work is to demonstrate that the integration of embedded processing, audio signal analysis and artificial intelligence (typically considered computationally demanding) can be effectively achieved on constrained hardware. In this context, Corellium proved to be a valuable tool for the early and intermediate stages of development and validation. The algorithms and processing flow were initially designed and simulated in MATLAB and Simulink, while the neural network responsible for sound recognition was developed and trained within the same environment. This preliminary stage was essential to define the system architecture, validate the overall logic and gain a deeper understanding of how to structure the subsequent embedded implementation. The complete system was then adapted to the embedded domain through the Corellium virtual board, enabling a smooth transition from model-based design to firmware development. This virtualization environment provided a faster and more flexible testing framework than physical hardware, allowing rapid iteration, early validation of real-time execution and the identification of integration issues well before deployment. Finally, the algorithms were tested and further refined on the physical STM32U5-based hardware, confirming their robustness and real-time performance. Moreover, the proposed approach can be adapted to wearable platforms, such as smart glasses or portable devices, providing a more intuitive and user-friendly way to signal potential hazards. In such implementations, visual and haptic notifications could be integrated to deliver discrete yet effective alerts, further enhancing environmental awareness and user safety for individuals with hearing impairments.