@article{LI2026, 
author = {Shanshan LI and Minglai YANG and Zhe ZHANG and Chuanheng SUN and Shanning WANG and Changyuan ZHAI and Wenyong LI},
title = {Flight trajectory tracking method for peach borer based on improved YOLOv12+Deep OC-Sort},
year = {2026},
journal = {Transactions of the Chinese Society of Agricultural Engineering},
volume = {42},
number = {11},
pages = {299-311},
keywords = {Conogethes punctiferalis, no visible light monitoring, multi target tracking, trajectory optimization, YOLO},
url = {https://www.sciopen.com/article/10.11975/j.issn.1002-6819.202601048},
doi = {10.11975/j.issn.1002-6819.202601048},
abstract = {Flight trajectory tracking of peach borer (Carposina sasakii) adults is confined to their small targets, strong motion nonlinearity, and frequent short-term occlusions in the absence of visible light. Peach borer adults typically exhibit rapid, irregular, and highly nonlinear flight behaviors in nocturnal or infrared imaging environments, with weak visual contrast, blurred contours, and intermittent target disappearance. Reliable detection and tracking often require long-term trajectory behavioral analysis due to frequent identity switches and severe trajectory fragmentation. It is also insufficient for the isolated optimization of either the detection or tracking module. A joint optimization strategy is also required to enhance the overall robustness of the trajectory acquisition. In this study, a detection–tracking collaborative framework was proposed to integrate target detection and multi-object tracking for robust, continuous, and identity-consistent trajectory acquisition. The detection was optimized to provide more stable observations, while the tracking was designed to better tolerate short-term detection failures and abrupt motion. In target detection, an improved detection architecture was constructed to enhance robustness against complex backgrounds, weak infrared contrast, and dense target distributions. Specifically, a C3k2_DEAB module was introduced into the backbone network to strengthen feature representation for small-scale flying targets. The effective local receptive field was expanded to incorporate attention-driven feature discrimination. The module improved the detector’s sensitivity to weak, fragmented, and low-contrast target cues in infrared imagery. In addition, a BiFPN multi-scale feature fusion was employed to adaptively integrate semantic and spatial information at different feature levels. Detection completeness was enhanced for the small targets with varying scales, motion states, and imaging conditions, thereby reducing missed detections caused by scale variation and rapid motion. Distance-IoU-based Non-Maximum Suppression (DIoU-NMS) was incorporated into the inference stage under target overlap and dense flight scenarios. According to the overlap and center distance between bounding boxes, DIoU-NMS effectively reduced false suppression in crowded scenes. Both localization accuracy and detection recall were improved for closely spaced flying insects. In the multi-object tracking, Kalman filter prediction optimization was incorporated with iterative motion smoothing under the nonlinear and abrupt motion patterns of peach borer adults. Iterative smoothing was introduced into the state prediction, indicating the tracker more resilient to sudden velocity and short-term observation noise in insect flight trajectories. Furthermore, a trajectory consistency constraint strategy with a predefined target number was designed to reduce the short-term missed detections. The constraint prevented premature trajectory termination and erroneous identity reassignment caused by transient detection failures or occlusions. A trajectory optimization and repair mechanism was developed to enhance tracking robustness, particularly for short-term trajectory fragmentation. Temporarily unmatched trajectories are maintained to continuously evaluate within a temporal window, according to motion consistency and spatial feasibility. Effective identity recovery and continuity maintenance were realized under short-term occlusion, various directions, and dense interactions, while avoiding erroneous long-term trajectories. Extensive experiments were conducted on infrared imaging in the laboratory. The improved model significantly outperformed the baseline YOLOv12 + Deep OC-Sort framework. Among them, AssA, IDF1, HOTA, MOTP, and MOTA were improved by 3.42, 3.74, 1.85, 2.22, and 3.85 percentage point, while the number of trajectory fragmentations (Frag) was reduced by 305 instances. Visualization results confirm that the framework effectively suppressed trajectory fragmentation and identity redistribution under high-speed motion and short-term occlusions, thus producing smoother, longer, and more coherent flight trajectories. Systematic and collaborative optimization of detection and tracking stages can be expected for accurate and reliable characterization of peach borer adult flight trajectories under non-visible-light conditions. The framework can also provide a solid data reference for subsequent quantitative analysis of flight behavior, phototactic response modeling, and the intelligent pest monitoring technologies in complex agricultural environments.}
}