Publication
Consolidation of symbolic instances using sensor data via tracklet merging for long-term monitoring of crops
Mark Niemeyer; Joachim Hertzberg; Grzegorz Cielniak
In: 9th Computer Vision in Plant Phenotyping and Agriculture. Computer Vision in Plant Phenotyping and Agriculture (CVPPA-2024), located at 18th European Conference on Computer Vision ECCV 2024, September 29-30, Milano, Italy, Springer, 2024.
Abstract
Long-term monitoring of plants is crucial for many practical applications in agriculture. This task requires associating the sensor data over long timeframes to a single individual object representing a unique symbolic instance. In general, this is a well-known tracking problem, but state-of-the-art trackers cannot track objects over their whole lifetime reliably, due to occlusions and changes in the object's appearance, thus only sub-tracks / tracklets are created. This paper proposes a methodology to consolidate the tracklets into the real track. For the consolidation of two tracklets matching costs using spatial, temporal and appearance-based information are calculated and the overall matching of the tracklets is optimised using the Hungarian algorithm. Using a strawberry tracking dataset for the evaluation, the presented methodology shows promising results and can match the tracklets with an accuracy of 72.5 %.