Work

FlyYOLO-SORT: A lightweight detection and multi-object tracking framework for tephritid fruit flies towards automated phytosanitary monitoring

Year: 2026

Type: article

Abstract: Real-time monitoring of tephritid fruit flies is essential for effective agricultural pest control and stringent plant quarantine measures. Most existing computer vision approaches primarily focus on static images derived from sticky traps. The absence of temporal motion information makes it difficult for these methods to maintain robust tracking performance when targets undergo fast and drastic movements, morphological blurring, and mutual occlusion. To overcome these limitations, mining and utilizing temporal motion features is critical, with high-precision multi-object tracking (MOT) serving as the foundation. Using Bactrocera dorsalis and Bactrocera correcta as case studies, we construct a high-performance detection and MOT framework designed for video monitoring. First, we propose FlyYOLO, a lightweight detector based on YOLOv11n. By fusing deep and shallow features, the detector is encouraged to focus on subtle dorsal features of the fruit flies, thereby improving classification accuracy. Simultaneously, a bounding box scoring mechanism is introduced to further optimize localization precision. Second, using FlyYOLO as the base detector, we propose FlyYOLO-SORT, an algorithm based on the tracking-by-detection (TBD) paradigm. We design a Dual-Box Complementary Fusion (DBCF) strategy to effectively integrate the strengths of Axis-Aligned Bounding Boxes and Oriented Bounding Boxes to generate high-quality detection results. Additionally, an Angle-Aware Euclidean Metric (AEM) is proposed for data association, significantly enhancing tracking robustness in complex scenarios involving frequent occlusion and irregular flight. On our dataset, FlyYOLO-SORT achieves outstanding performance with 48.26% HOTA, 80.15% MOTA, 56.01% MOTP, and 81.96% IDF1. Compared to ByteTrack, it reduces ID switches by 52.3%. Experimental results demonstrate that the proposed method enables efficient and stable MOT of fruit flies, providing reliable technical support for applications such as population counting and behavioral analysis of insects exhibiting complex motion patterns.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2026.102177