StableSort-CMC: A Tracking Algorithm for Robot Dog Based Orchard Fruit Counting

In recent years, the utilization of deep learning technology for fruit yield prediction methods using robot dogs equipped with cameras, along with target detection and tracking algorithms, has emerged as a research focus. However, the movement of robot dogs through undulating orchard terrain often induces vibration in the mounted cameras, which can significantly impact the performance of tracking algorithms. To address this issue, this paper proposes a tracking algorithm integrated with a camera motion compensation module. This module applies image rotation, scaling, and translation parameters to the tracking algorithm, enhancing its performance and counting accuracy in vibrated scenes. Utilizing nectarine data collected by a robot dog equipped with a gimbal camera, the proposed algorithm achieves a Multiple Object Tracking Accuracy (MOTA) of 63.4%. In ablation experiments, integrating the camera motion compensation module into the tracking algorithm improves MOTA by 2.9%. These results validate the efficacy of the proposed algorithm in mitigating the effects of camera vibration and enhancing tracking performance in dynamic orchard environments