Sidewalk weed detection is crucial for maintaining urban environmental quality and ensuring residents' comfort, yet existing object detection methods have significant research gaps in this task. First, complex sidewalk backgrounds with variable tiles, shadows, and fallen leaves cause weed-background feature confusion, while conventional feature extraction lacks effective interference suppression. Second, traditional upsampling modules fail to balance computational efficiency and reconstruction accuracy, hindering real-time edge deployment on weeding robots. Third, existing loss functions poorly address weeds' multi-scale characteristics and irregular morphologies, with insufficient small-target weights, ignored directional offsets, and coupled width-height optimization leading to unstable regression. To solve these issues, this paper proposes Sidewalk Weed - You Only Look Once (WayWeed-YOLO) based on You Only Look Once version 13 (YOLOv13): we propose a Weed-Aware Attention (WAA) Mechanism that integrates channel and spatial attention to enhance weed features and suppress backgrounds; the ultra-lightweight Dynamic Sample-based Upsampler (DySample) reduces parameters for edge deployment; we propose a Weed-Intersection over Union (Weed-IoU) loss function incorporating small-target compensation, directional anti-offset, and independent width-height optimization to improve multi-scale detection. Experiments show WayWeed-YOLO outperforms state-of-the-art models, including Faster Region-based Convolutional Neural Network (Faster R-CNN), Single Shot MultiBox Detector (SSD), and YOLO series, with 3.3% higher precision than original YOLOv13 and 42.7 Frames Per Second (FPS) on NVIDIA Jetson Orin NX, balancing accuracy and real-time performance. A dedicated visual interface enables real-time monitoring. Its novelty lies in targeted solutions to sidewalk weed detection challenges, providing reliable algorithmic support for intelligent weeding robot deployment on edge devices.
Dingran Wang (Sat,) studied this question.