<p>The accurate identification of non-cooperative space targets is challenged by the time-varying dynamics inherent to orbital maneuvers. Conventional holistic classification methods often assume constant dynamics, resulting in performance degradation when analyzing hybrid trajectories with evolving intents. To address this, we propose the Trajectory Multi-stage Processing and Classification (TMPC) framework, which adopts a “decomposition–reconstruction” strategy. First, we introduce a singularity-based segmentation mechanism that utilizes extrema in the relative range profile to decouple hybrid trajectories into dynamically consistent primitives. Second, we employ an MDL-based feature extraction method acting as a structural filter, which effectively isolates the essential trajectory signal from stochastic measurement noise. Finally, a density-adaptive hierarchical classifier routes sub-trajectories to specialized recognition modules based on the density distribution in the feature space. Experimental validation demonstrates that the framework achieves a classification accuracy exceeding 97% for single-intent trajectories and robustly identifies evolving intents in complex hybrid scenarios. Notably, the framework exhibits superior noise robustness, maintaining high performance under low signal-to-noise ratio conditions, while achieving an average data compression ratio of 72.43%.</p>