A Multi-Scale Feature Fusion Method Based on DenseNet for UWB Radar Gesture Recognition

This project presents a gesture recognition system using ultra‑wideband (UWB) radar sensors and deep learning. The dataset contains range‑time maps from three radar sensors placed at left, right, and top positions, covering twelve dynamic hand gestures. A baseline multi‑scale feature fusion DenseNet (MSF‑DenseNet) is implemented, which uses parallel convolutions of different kernel sizes to extract spatial features. Two novel models are proposed to improve recognition capability. The first, DMSF‑DenseNet, introduces dynamic weighting of the multi‑scale convolution branches based on input content. The second, CSA‑DenseNet, adds a cross‑sensor attention module that learns to reweight the three radar channels before feature extraction. The models are trained, validated, and tested using a fixed data split. A web application is developed using Flask, HTML, CSS, and JavaScript, providing modules for user registration, login, image upload, classification, and logout. The best performing model is deployed for inference, returning predicted gesture labels with confidence scores. The system demonstrates that attention‑based sensor fusion enhances radar‑based gesture recognition without adding computational overhead.

Keywords: UWB radar, gesture recognition, DenseNet, multi‑scale feature fusion, cross‑sensor attention, dynamic convolution, range‑time maps, deep learning, Flask web application, sensor fusion.