In multi-domain collaborative sensing and target situation awareness tasks， traditional high-accuracy Direction-of-Arrival （DOA） estimation methods typically rely on a known and fixed number of sources. However， during continuous tracking， the number of targets may vary over time， causing fixed-source-number models to become invalid. Meanwhile， in passive array sensing， multi-target signals often appear in a superimposed form， making it difficult to construct effective single-target likelihood functions and thus limiting the applicability of Random Finite Set （RFS）-based filters in DOA tracking. To address these challenges， this paper proposes a continuous DOA tracking method based on a Taylor-expanded projection-cancellation model. By performing a Taylor expansion of the steering vector around the predicted angle， an extended signal subspace is constructed， and interference from other sources is eliminated using sequential projection operators， resulting in a single-source-equivalent pseudo-spectrum. This enables robust decomposition of multi-source superimposed observations and reliable construction of single-target likelihoods. Furthermore， based on the energy characteristics of the projection-cancellation spectrum and the particle-weight distribution， a birth-death detection mechanism is introduced， which adaptively adjusts the likelihood dimension in the PHD filter without requiring prior source-number estimation. Simulation results demonstrate that the proposed method achieves superior DOA tracking performance compared with conventional approaches， particularly under low SNR conditions and dynamically varying source scenarios.