چكيده لاتين
Multiple Object Tracking is a vital component in intelligent navigation systems and perception-based technologies, playing a central role in applications such as autonomous vehicles, mobile robots, and advanced environmental monitoring systems. A significant portion of existing state-of-the-art methods is designed based on simple and deterministic motion models, such as constant velocity and constant acceleration models, and performs tracking solely by relying on a single prediction of each object’s future state. While these approaches, due to their structural simplicity and high computational efficiency, offer acceptable performance in many conventional tracking scenarios, they experience a notable decline in tracking accuracy and stability when confronted with the inherent complexities of dense and uncertain environments, including high object velocities, abrupt maneuvers, intermittent reappearances, and partial or full occlusions. To address these challenges, this dissertation proposes a novel method, termed PMM-MOT, aimed at overcoming the limitations of conventional single-hypothesis estimation approaches by adopting a multi-hypothesis strategy. In this approach, for each active track, a set of plausible motion predictions is generated simultaneously, and the probability of each prediction is independently estimated based on a data-driven analysis of the object’s recent motion behavior. The key innovation of PMM-MOT lies in producing multiple predictions for each active track, predictions that are not generated randomly or independently of the object’s dynamic state, but rather through an adaptive analysis that leverages feedback from the recent motion history and estimates the deviation angle relative to the current heading. In the first stage, the deviation angle is estimated using a deep learning model based on a GRU network. Subsequently, guided by this estimate, alternative motion trajectories are generated in both left-turning and right-turning directions relative to the reference path, enabling a targeted representation of the object’s maneuvering space in accordance with its behavioral characteristics. Furthermore, to evaluate the physical plausibility of each predicted trajectory, the proposed method employs a logistic regression model based on the physical quantity of momentum to estimate the likelihood of each prediction in proportion to the object’s dynamic properties. In other words, trajectories that exhibit greater consistency with the object’s inertia and physical characteristics are assigned higher occurrence probabilities and are consequently more likely to achieve successful alignment with new detections during the data association process. To comprehensively assess the performance of PMM-MOT, a series of experimental evaluations was conducted on the widely recognized KITTI and nuScenes benchmark datasets. In the 2D MOT evaluation on KITTI, PMM-MOT achieved HOTA scores of 80.27 for the Car class and 52.48 for the Pedestrian class, along with high AssA and MOTP values, outperforming other notable approaches. Moreover, in the 3D MOT evaluation on nuScenes, the proposed method attained the highest reported AMOTA score of 75.5 while significantly reducing the number of identity switches, ranking first among all reported methods.
