Improvnig MOT Based on Optimization of Learnable Kalman Filter: Trajectory Visualization-Based Data Augmentation and Loss Function Development

In this paper, we propose a novel optimization method for a learnable kalman filter to enhance multi-object tracking system by integrating an optimized loss function with a learnable Kalman Filter and applying image data augmentation techniques to trajectory data which type of time-series data. The way of using traditional Kalman Filters often fail to adapt to dynamic changes inherent in MOT scenarios. By incorporating deep learning directly into the filter’s estimation process, proposed method provides a more flexible and accurate tracking mechanism. Our experiments on standard benchmarks such as MOT17 and MOT20 demonstrate improvements in association accuracy, and location estimation, particularly highlighting the balanced performance across various metrics including HOTA, MOTP and FP. Novelty of this paper presents a unified framework combining MOT system with learnable Kalman filter. The results confirm that utilizing a tailored loss function alongside sophisticated data augmentation can elevate MOT performance. We anticipate that our findings will spur further research into adaptive filtering techniques within MOT applications.