Multi-object tracking for apple counting in orchards using stereo vision

Quantitative fruit counting at the individual tree level is a fundamental requirement in data-driven precision agriculture and a critical input for yield estimation models. However, achieving consistent tracking and accurate detection of fruits in unstructured outdoor environments remains a challenging task due to factors such as occlusion, varying illumination and sensor noise. Moreover, conventional 2D image-based methods are prone to scale ambiguity and duplicate detections from multiple viewpoints. To overcome these challenges, this work presents a 3D fruit counting framework based on stereo cameras in which each detection is localized within a georeferenced spatial coordinate system using GPS and visual odometry. The fruits are detected using a YOLOv8x model to identify object positions within individual images. To enable tracking across multiple consecutive frames, two state-of-the-art algorithms, OC-SORT and ByteTrack, are benchmarked against each other. OC-SORT, with incorporated optical flow, achieved the best performance on the validation dataset, with HOTA = 0.59, MOTA = 0.52, and IDF1 = 0.67, while the detection model reached a mAP<inf>0.5:0.95</inf> = 0.56. The proposed fruit counting framework demonstrated promising performance on Elstar trees, yielding an estimated total counting error of 20 apples, corresponding to a RMSE of 6.24 % and an R<sup>2</sup> score of 0.96. Given that fruit count is a key parameter to yield prediction, these results support accurate yield mapping and informed precision orchard management.

Subjects

3D-Mapping

Fruit Counting

Multi-Object Tracking

Precision Agriculture

Visual Odometry

Yield Estimation

DDC Class

600: Technology

Options

Multi-object tracking for apple counting in orchards using stereo vision