CropTrack: A Tracking with Re-Identification Framework for Precision Agriculture

TL;DR

CropTrack tackles the challenging problem of multi-object tracking in agricultural settings by fusing appearance-based re-identification with motion cues. It introduces a reranking-enhanced appearance association, a one-to-many matching scheme with greedy conflict resolution, and an EMA-based prototype feature bank, powered by an NSA Kalman filter for robust state estimation. On TexCot22 and AgriSORT-Grapes, CropTrack achieves state-of-the-art identity preservation and competitive overall tracking metrics, while reducing identity switches and maintaining track continuity under occlusion and visual similarity. The method offers practical impact for precision agriculture by enabling reliable crop and fruit monitoring for targeted interventions and yield estimation.

Abstract

Multiple-object tracking (MOT) in agricultural environments presents major challenges due to repetitive patterns, similar object appearances, sudden illumination changes, and frequent occlusions. Contemporary trackers in this domain rely on the motion of objects rather than appearance for association. Nevertheless, they struggle to maintain object identities when targets undergo frequent and strong occlusions. The high similarity of object appearances makes integrating appearance-based association nontrivial for agricultural scenarios. To solve this problem we propose CropTrack, a novel MOT framework based on the combination of appearance and motion information. CropTrack integrates a reranking-enhanced appearance association, a one-to-many association with appearance-based conflict resolution strategy, and an exponential moving average prototype feature bank to improve appearance-based association. Evaluated on publicly available agricultural MOT datasets, CropTrack demonstrates consistent identity preservation, outperforming traditional motion-based tracking methods. Compared to the state of the art, CropTrack achieves significant gains in identification F1 and association accuracy scores with a lower number of identity switches.

Figures (5)

• Figure 1: A comparison of state-of-the-art trackers on the AgriSORT-Grapes saraceni2024agrisort dataset. The horizontal axis is the identification F1 (IDF1) score, the vertical axis is the association accuracy (AssA), and the radius of each circle corresponds to the higher-order tracking accuracy (HOTA) score. CropTrack achieves the best AssA score and comparable IDF1 and HOTA performance.
• Figure 2: An overview of the CropTrack pipeline where our key contributions are shaded in blue. The pipeline begins with an object detector that generates both low- and high-score detections. The high-score detections are processed by the first appearance-based association. Then, all unmatched detections and tracklets proceed to the second appearance-based association step. Finally, unmatched tracklets are processed with the low-score detections in the third IoU-based association.
• Figure 3: An example of IoU-based association failure. Left: In frame 16, a new track is initiated with bounding box ID 17. Right: By frame 43, the track (white bounding box, ID 17) has drifted from the object's true position, leaving the yellow detection box unmatched and failing to associate with the correct track.
• Figure 4: A qualitative comparison of (a) CropTrack, (b) NTrack, and (c) PineSORT on a test sequence from TexCot22 T8/5M9NCI_2024 across multiple time steps. The top row displays the tracking results at frame 47, while the middle and bottom rows show the corresponding results at frame 58 and 69, respectively. Each bounding box represents a tracklet and the color signifies its object ID. A single color is used for the same object, while different colors are employed for distinct objects. CropTrack yields superior ID preservation under strong occlusions.
• Figure 5: Tracking performance under varying detection accuracies.