The analysis of complex flight patterns and collective behaviors in swarming insects has emerged as a significant focus across biological and computational fields. Tracking these insects, like fruit fly, presents persistent challenges due to their rapid motion patterns and frequent occlusions in densely populated environments. To address these challenges, we propose a tracking method using particle filter framework combined with a Kolmogorov-Arnold Network (KAN)-Transformer model to extract the global features and fine-grained features of the trajectory. Additionally, manually annotated ground truth datasets are established to enable thorough assessment of tracking methods. Experimental results demonstrate the effectiveness and robustness of our proposed tracking method. Analysis of tracked trajectories revealed the Reynolds rules of flocking behavior.