Cluster tracking algorithms for autonomous driving

In this master’s thesis, an approach to tracking based on position and velocity estimates for Advanced Driver Assistance Systems (ADAS) is presented. The method enables basic vehicle motion tracking. The work begins from sensor-derived coordinates, which are inherently affected by noise, and addresses these measurement errors through appropriate processing and estimation techniques. To meet these requirements, a combination of complementary techniques is employed: the Kalman filter to reduce noise and provide accurate state estimates, the DBSCAN clustering algorithm to identify correlated point groups, and the Auction algorithm to optimize the assignment between existing tracks and new observations. The Auction algorithm, an efficient alternative to the Hungarian Algorithm, minimizes assignment costs while ensuring real-time performance. The system optimizes data-to-track association and produces experimental results that demonstrate good performance in terms of accuracy and computational speed. The research begins with synthetic datasets deliberately characterized by noise and then progresses to more realistic datasets generated using MATLAB’s Automated Driving Toolbox, which enables results closer to those obtainable under real operational conditions.