This paper presents the first edge computing framework for RTI systems combining intermediate sensor fusion with model quantization. Traditional RTI approaches [1] rely on centralized architectures (450 ms latency), while our key innovations enable 73% faster edge processing (120 ms) through hybrid ResNet architecture compressed via 8-bit quantization and intermediate fusion of RTI attenuation maps and RGB features. Our work introduces an on-device processing pipeline implemented on low-power IoT devices---including Jetson Nano, Raspberry Pi, and ESP32---that enables real-time inference without cloud-based computation.

This paper presents a system that fuses radio tomography data with RGB camera information for enhanced multi-person detection and tracking in indoor environments. Experiments were conducted in an irregularly shaped room with four subjects. Due to its limited resolution, radio tomography initially represented all individuals as a single entity. By incorporating RGB camera data, we were able to accurately identify and track each person individually. Our findings underscore the significance of integrating data from both modalities for improved detection and tracking performance in indoor surveillance and security systems.

Artificial Intelligence offers cost-effective solutions to improve business processes and ensure more satisfying customer service. The advantage of solutions based on artificial intelligence is the possibility of using the API with mobile or stationary applications and cloud services. The research presented here aims to develop a deep learning model using cross-modal techniques on the example of a multi-tasking network. The main task is to use computer vision on IoT devices using sensors for customer service. Additionally, the solution will be based on distributed systems. Finally, the method of building the multi-tasking model, which will be designed to determine the person in the image and their emotional state, will be verified.

The research aims to develop and optimize a deep learning model for faster inference using multi-task learning for embedded systems. Experiments using face photos and sound in the form of a spectrogram were prepared to verify the model's performance in recognizing a person and their emotional state. Research has shown that in IoT devices, the inference is faster when a multi-tasking model is used compared to a system based on several models, each responsible for inferring one thing.