VTC-TrackBot: A Robust Spatiotemporal Framework for Real-Time Chrysanthemum Counting on Mobile Edge Devices

Abstract

Accurate and real-time flower counting is a critical component of digital horticultural management, enabling precise yield estimation and harvest planning. However, applying computer vision in mobile field scenarios remains challenging due to dense cluster occlusion, variable illumination, and severe camera jitter caused by crawler platform vibrations. To address these issues, this article proposes VTC-TrackBot, an integrated edge-computing system for robust chrysanthemum counting. The framework reformulates the counting task as a reliable tracking-by-detection problem through three synergistic innovations. First, we develop an enhanced detector, you only look once–varieties of tea chrysanthemum, which integrates MSPPFCSPC and BoT3 modules to improve the detection of small and heavily occluded flowers, achieving a mean average precision of 86.6%. Second, to overcome identity switching induced by nonlinear platform vibration, we propose an improved observation-centric simple online and realtime tracking algorithm. Diverging from computationally intensive appearance-based methods, our approach incorporates a noise scale adaptive Kalman filter and a soft-weighted observation-centric momentum strategy to maintain robust trajectory consistency under vibration. Third, a dual-line spatiotemporal integration counting strategy is introduced to repair fragmented tracklets across the central blind zone and filter vibration-induced outliers. Extensive field experiments on a custom crawler robot equipped with an NVIDIA Jetson Xavier NX demonstrate that VTC-TrackBot operates at 23 frames per second, with a counting coefficient of determination ($R^{2}$) of 0.983 and a mean relative error of only 4.14%. These results demonstrate that the proposed lightweight spatiotemporal framework provides an effective and practical solution for real-time phenotyping and yield estimation of densely clustered crops in complex agricultural environments.