A Novel Data Association Algorithm for Tracking and Localization Using Segmented Data Analysis

Multiple object tracking is used in a range of applications, including autonomous control of various vehicles and sports analytics. The main challenges in developing tracking systems are detection errors and computational load that scales with the number of tracked objects. Therefore, the problem remains highly relevant. This paper formalizes the generalized multisensor tracking task for multiple objects. To address this task, a novel algorithm for object tracking and localization is proposed. It combines cluster analysis, relevance estimation for object data, and a set of Kalman filters that can be computed independently and in parallel. The algorithm operates at the data association stage and uses segmented data on object positions and additional features. The architectural design ensures compatibility with both modern methods and traditional approaches for detection and segmentation. The approach is suitable for a wide range of applications. The algorithm was tested on the publicly available APIDIS dataset, allowing direct comparison with existing methods and demonstrating competitive performance with a MOTA score of 81.8%. The feasibility of the algorithm for thermal updraft localization was demonstrated by tracking soaring objects using a custom dataset of videos recorded by a paragliding pilot. The results confirm the potential of this approach for deployment on autonomous unmanned gliders.