Radar incoherent interference can cause an increase in noise floor and generate false plots. This exacerbates data association ambiguity in multi-extended target tracking, leading to track fragmentation and identity confusion, consequently causing errors in target state and shape estimation. Therefore, this paper proposes a joint spatio-temporal clustering and probability hypothesis density (PHD) based radar anti-interference method for multi-extended target tracking. Firstly, addressing the random time-varying characteristics of both the number and locations of extended target scattering points, the random finite set (RFS) theory is employed to model the multi-extended target state and measurement sets. High-quality partitioning of the dynamic, time-varying measurement sets is achieved by integrating spatio-temporal clustering. This approach not only avoids the complex explicit data association operations introduced by false targets but also resolves the partition explosion problem arising from increased measurement set dimensionality. Furthermore, by leveraging the probability hypothesis density (PHD) function, the interference from false targets is eliminated through the weighted summation of the Gaussian distributions and inverse Wishart distributions corresponding to different measurement subsets. This enables precise tracking of both the motion trajectories and shapes of multiple extended targets. Finally, experimental results under scenarios involving crossing trajectories, varying signal-to-noise ratios (SNR), and variable target numbers demonstrate that the proposed method achieves tracking points with optimal sub-pattern assignment (OSPA) distance below 0.5 m for over 95% of instances. It outperforms the K-means clustering-based PHD filter, the extended target generalized inverse Wishart PHD (ET-GIW-PHD) filter, and random matrix-based methods.