CV

PhD Researcher in Computer Vision

AImageLab, University of Modena and Reggio Emilia

Research on 3D scene and human reconstruction using NeRF and Gaussian Splatting. Model compression and efficiency enhancement for real-time inference on edge devices. Emergency response and surveillance classification models for drones and UAVs.

• Efficient Deep Learning
• CNN Architecture Design
• Model Compression & Quantization
• AI Deployment on Edge Devices
• 3D Human and Scene Reconstruction

Robotics Research Intern in Computer Vision

Toyota Motor Europe

Designed transformer-based models for 3D human pose estimation and temporal tracking. Developed CNN-based ergonomics monitoring system with web dashboard.

• 2D Hand Pose Estimation
• CNNs for 3D Human Pose Estimation
• Deep Human Ergonomics Estimation and Analysis
• Transformers for 3D Human Pose Estimation and Tracking

Computer Vision & Firmware Engineer

ProjectRED

Implemented real-time object detection for robotic pick-and-place tasks using YOLOv5 and ROS. Developed ArUco-based navigation and CAN bus firmware for STM32.

• ROS
• STM32
• CAN Bus
• ArUco Marker Detection
• YOLOv5 Object Detection

Software & Firmware R&D Engineer

WATER LINE S.R.L.

Developed firmware in C/C++ for IoT drink dispensers. Built desktop tools with C# WPF and PyQt5.

• Embedded Systems
• STM32 and ESP32
• C/C++ Keil MDK
• Python PyQt5
• C# WPF

PhD

University of Modena and Reggio Emilia

Computer Vision

• Advanced Methodologies and Technologies for Neurophysiological Research
• Machine Learning Applications in Trading and Portfolio Management
• Explainable AI in the Neural Network Domain
• Computer Vision for Spatial Intelligence (ICVSS 2025)

M.Sc.

University of Modena and Reggio Emilia

Artificial Intelligence Engineering

• Computer Vision
• IoT & 3D Intelligent Systems
• Smart Robotics
• Reinforcement Learning
• Distributed AI
• Power Electronics

B.Sc.

University of Modena and Reggio Emilia

Computer Engineering

• Computer Architecture
• Calculus I & II, Multivariable Calculus
• Data Structures and Algorithms
• Operating Systems
• Telecommunications
• Digital Electronics

International Computer Vision Summer School (ICVSS)

University of Catania, Italy

Selected participant for advanced lectures on Computer Vision, 3D Scene Understanding, and Efficient Deep Learning. Successfully passed the final examination.

TakuNet: Energy-Efficient Models for Real-Time Aerial Disaster Response and Monitoring on Edge Devices

Image and Vision Computing, vol. 175, art. 106151 (Elsevier)

Journal extension of the WACVW 2025 paper, with Gianluca Filippini, Andrea Torlai, Guido Borghi and Roberto Vezzani. Introduces TakuNetV2, a redesign that restores cross-channel mixing inside the blocks, and measures the family on five embedded platforms, from CPUs to the Hailo-8 NPU.

BoltNet: An Ultra-Lightweight Convolutional Network for On-Device Plant Species Identification

CVPPA Workshop at the European Conference on Computer Vision (ECCV)

0.682 F1 with 341K parameters (1.37 MB) on Pl@ntNet-300K, the best score of any model under 2 MB, and the highest frames per second per watt on both the Hailo-8 NPU and the Jetson Orin Nano.

AutoDock-AI: a Docker-based Framework Accelerating Artificial Intelligence Research

GitHub (Software)

Docker-based framework providing automatic device detection, environment setup, and seamless development workflows for AI research on embedded platforms.

TakuNet: an Energy-Efficient CNN for Real-Time Inference on Embedded UAV systems in Emergency Response Scenarios

IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)

Presented an energy-efficient CNN designed for real-time inference on embedded UAV systems. Achieves near state-of-the-art accuracy with extreme efficiency on edge devices like Jetson Orin Nano and Raspberry Pi

BoltNet

Ultra-lightweight fully convolutional network for fine-grained, high-cardinality classification on constrained hardware. Two parameter-free rearrangements trade channels for spatial resolution.

• CVPPA at ECCV 2026
• 341K parameters, 1.37 MB
• Best F1 under 2 MB on Pl@ntNet-300K
• 2,354 FPS/W on Hailo-8

TakuNetV2

Journal extension of TakuNet, a family of ultra-lightweight networks for real-time aerial disaster response under 100K parameters, benchmarked on five embedded platforms.

• Image and Vision Computing 2026
• 0.959 F1 on AIDERv2, 0.838 on CLRS
• 560 FPS on Jetson Orin Nano
• Evaluated on the Astrial Hailo-8 board

OpenNPU

From-scratch educational build of a Neural Processing Unit on FPGA, documenting the path from systolic array to compiled inference.

• Verilog RTL
• Systolic array
• Educational, built from zero

EdgePowerMeter

High-precision power monitoring system for edge AI devices with real-time visualization and analysis tools.

• Energy (V,A,W) Monitoring and Analysis
• Harmonic Analyzer
• Real-time Processing

TakuNet

Energy-efficient CNN for real-time inference on embedded UAV systems in emergency response scenarios. Achieves near-SOTA accuracy with minimal computational requirements.

• WACVW 2025
• ONNX & TensorRT Support
• 650+ FPS on Jetson Orin Nano
• 50+ GitHub stars

AutoDock

Docker-based framework accelerating AI research with automatic device detection and environment setup for Jetson and Raspberry Pi platforms.

• Docker-based framework
• Simple build and run
• Automatic device detection
• Seamless development workflow

RL-Swarm

Multi-agent reinforcement learning framework with pheromone-based perception for swarm intelligence applications.

• SARSA
• DQN
• Centralized vs Decentralized
• Supported Slime Env development

Multi-Vision System for Autonomous Driving

Computer Vision Framework for Vehicle Detection and Tracking with YOLOv7 and Kalman Filters, Lane Assistant and Traffic Sign Recognition.

• YOLOv7 Training on BDD100K (5xRTX3090 Grant)
• Kalman Filters tracking
• Computer Vision algorithms
• Inference on Jetson Nano (OpenCV CUDA)