FA-YOLO: Adaptive Feature Fusion and Multi-Scale Dilated Attention for Fall Detection

In this paper, we present FA-YOLO (Adaptive Feature Fusion and Multi-Scale Dilated Attention YOLO) which is designed to enhance collision detection performance. Our approach integrates advanced techniques to improve feature extraction and fusion. We employ Adaptive Feature Fusion (ASFF) to seamlessly merge multi-scale feature maps, allowing the model to utilize both fine and coarse details. The BiFormer attention mechanism, specifically BiLevelRoutingAttention, is incorporated to bolster the model's ability to capture global context and local nuances. Additionally, a Multi-Scale Dilated Attention (MSDA) module is introduced to further refine feature representation by capturing contextual information across multiple scales using dilated convolutions. To evaluate FA-YOLO, we created a custom Collision Detection Dataset (CDD) comprising 1466 images from various online sources, annotated with category labels and bounding boxes using the LabelImg tool. The dataset is split into 1133 images for training and 333 images for validation. Experimental results demonstrate that FA-YOLO achieves a Mean Average Precision (MAP) of 60.86% on the CDD, significantly outperforming the baseline YOLOv5 model, which achieves a MAP of 53.21<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, representing a 14.3% improvement. These results highlight the effectiveness of our proposed enhancements in collision detection tasks, showcasing the potential of FA-YOLO for applications requiring precise and reliable detection capabilities.