WeedsSORT: A weed tracking-by-detection framework for laser weeding applications within precision agriculture

Abstract

In precision agriculture, the application of artificial intelligence and high-power laser technology for weed control offers significant efficiency and accuracy advantages. However, it still encounters numerous challenges in the detection and tracking of weed targets. In terms of object detection, the variability in the size and specifications of weeds can result in the missed detection of smaller weed targets. Regarding tracking prediction, the similarity in weed shapes may result in reduced pose estimation accuracy, and the random motion of cameras within laser weeding systems further increases the risk of tracking failures. To address these challenges, this study introduces a spatial attention mechanism to enhance weed detection accuracy. It employs optimized multi-feature layer extraction and optimal feature matching algorithms to derive motion estimation results. Ultimately, an adaptive extended Kalman filtering algorithm is integrated to establish a weed tracking algorithm that correlates temporal and spatial information, ultimately achieving rapid and precise detection and tracking of weeds in laser weeding scenarios. The detection accuracy of the optimized algorithm was tested on both publicly available datasets and self-collected detection datasets, achieving a mean Average Precision (mAP) of 97.29% and 85.83%, respectively. Furthermore, tracking performance was evaluated using the LettuceMOT dataset and the self-collected WeedsMOT dataset, demonstrating improvements in Higher-Order Tracking Accuracy (HOTA) accuracy of 12.01% and 8.75% when compared to the ByteTrack and DeepOCSORT algorithms. The experimental findings substantiate the efficacy of the proposed weed detection and tracking algorithm, offering a valuable reference for the progression of laser weeding technology within precision agriculture.